Elicitation non-paramétrique de la fonction d’utilité et de

l’aversion aux pertes sous l’hypothèse ”prospect theory”

Hilda Kammoun

To cite this version:

Hilda Kammoun. Elicitation non-paramétrique de la fonction d’utilité et de l’aversion aux pertes sous
l’hypothèse ”prospect theory”. Sciences de l’Homme et Société. Arts et Métiers ParisTech, 2007. Français.
⟨NNT : 2007ENAM0025⟩. ⟨pastel-00003225⟩

HAL Id: pastel-00003225

[Link]
Submitted on 20 Dec 2007

HAL is a multi-disciplinary open access archive L’archive ouverte pluridisciplinaire HAL, est des-
for the deposit and dissemination of scientific re- tinée au dépôt et à la diffusion de documents scien-
search documents, whether they are published or not. tifiques de niveau recherche, publiés ou non, émanant
The documents may come from teaching and research des établissements d’enseignement et de recherche
institutions in France or abroad, or from public or pri- français ou étrangers, des laboratoires publics ou
vate research centers. privés.

HAL Authorization
 ECOLE NATIONALE SUPERIEURE D'ARTS ET METIERS
DE PARIS

Ecole Doctorale « Sciences de la Décision et de l’Organisation » - ED 471

Equipe de Recherche GRID - UMR CNRS 8534

ELICITATION NON-PARAMETRIQUE DE LA FONCTION D’UTILITE ET DE

L’AVERSION AUX PERTES SOUS L’HYPOTHESE ‹‹PROSPECT THEORY››

THESE
présentée et soutenue publiquement le 28 septembre 2007
en vue de l’obtention du
DOCTORAT EN SCIENCES ECONOMIQUES
par

Hilda KAMMOUN
JURY

Directeur de thèse : Monsieur Mohammed Abdellaoui

Directeur de Recherche CNRS
GRID, ENSAM-ESTP

Rapporteurs : Monsieur Han Bleichrodt

Professeur
Université Erasmus de Rotterdam
Monsieur Philippe Delquié
Professeur
INSEAD
Suffragants : Madame Nathalie Etchart-Vincent
Chargée de Recherche au CNRS
CIRED

Monsieur Olivier L’Haridon

Maître de Conférences
IUFM de Paris, GRID
ELICITATION NON-PARAMETRIQUE DE LA FONCTION D’UTILITE ET DE
L’AVERSION AUX PERTES SOUS L’HYPOTHESE ‹‹ PROSPECT THEORY››

Résumé de la thèse : Dans ce travail, les fonctions d’utilité des gérants de portefeuilles sont
élicitées et leurs degrés d’aversion aux pertes mesurés sous l’hypothèse de la théorie des
prospects (1992) et suivant la méthode non-paramétrique d’Abdellaoui et al. (2006). Les
résultats obtenus sur le terrain corroborent les résultats obtenus par ces derniers au laboratoire
quant à la concavité de la fonction d’utilité pour les gains et la convexité pour les pertes. En
ce qui concerne l’aversion aux pertes, nos observations confirment son existence; néanmoins,
le gérant de portefeuilles médian est moins averse aux pertes que l’étudiant médian. Les
conditions qui caractérisent une expérience réelle du marché mais qui sont difficiles à
reproduire dans le contexte artificiel du laboratoire pourraient expliquer les différences de
comportement: notamment, la volatilité du marché boursier, les compensations incitatives de
Wall Street et le fait que les gérants de portefeuilles acquièrent sur le terrain une gamme de
formation et un haut niveau de connaissance qui font qu’ils évaluent les enjeux différenmment
des étudiants. La fonction d’utilité doit néanmoins, refléter les préférences de l’individu et
l’utilité ne doit pas changer selon la méthode utilisée. En effet, l’étude qualitative des
préférences d’étudiants en MBA suivant la méthode non-paramétrique de Baucells et
Heukamp (2006) confirme les résultats d’Abdellaoui et al. (2006) pour étudiants. Il est à noter
cependant que les étudiants changent de préférence (ne sont plus averses aux pertes mais
recherchent le gain) quand l’une des deux loteries offre une plus grande probabilité globale de
gain ou une plus grande probabilité de gain maximal combinée avec une perte extrême
limitée.

Mots clés : théorie des prospects, aversion aux pertes, utilité pour les gains et pour les pertes,
élicitation de mi-points, dominance stochastique de deuxième ordre, fonction de
transformation des probabilités.

PARAMETER-FREE MEASUREMENT OF THE UTILITY FUNCTION AND LOSS

AVERSION UNDER PROSPECT THEORY

Abstract : This work elicits the utility functions of financial practitioners and measures their
loss aversion coefficients under prospect theory (1992) using the parameter-free method of
Abdellaoui et al. (2006). The measurements in the field corroborate the latter’s measurements
in the laboratory regarding the concavity of the utility function for gains and convexity for
losses. However, although loss aversion exists in the aggregate, the median practitioner is
found to be less loss averse than the median student. Conditions that characterize a real
market experience but are difficult to realize in the artificial context of the laboratory may
account for the behavioral difference. Among them are the schooling in the assessment of
prospects, the volatility of the market and the Wall Street’s compensation incentives. An
important proviso is that the preferences of the students/practitioners analyzed following
another method reflect consistent preferences. The qualitative investigation of the preferences
of MBA students using the parameter-free method developed by Baucells and Heukamp
(2006) supports the results of Abdellaoui et al.’s (2006) for students. A noteworthy result is
the strong tendency to shift from loss aversion to gain seeking for the higher overall
probability of gain or the higher probability of maximal gain combined with a limited extreme
loss.

Keywords : prospect theory, loss aversion, utility for gains and losses, elicitation of
midpoints, second order stochastic dominance, probability weighting function.
L'ENSAM Paris n’entend donner aucune approbation ni improbation aux opinions émises

dans les thèses; ces opinions doivent être considérées comme propres à leurs auteurs.

1
ACKNOWLEDGMENT

I would like to express my sincere gratitude to Professor Mohammed ABDELLAOUI for

having followed the elaboration of this research up close and for having provided me
with the software he has designed without which it would have been difficult to conduct
an experimental work in the stock market. I would like also to thank him for his
availability, his thorough guidance as well as his clear and concise explanations despite
the geographical distance. I would like also to thank Professor Bertrand MUNIER for
having welcomed me at GRID, with much kindness and support.

Special thanks are also due to the members of my family in the U.S. the KURI and to my
relatives and friends in Lebanon who have personally contacted the financial practitioners
for the interviews. I am grateful to all the portfolio managers and financial advisors who
have graciously accepted my request for their time and who have cooperatively
responded during the interviews sharing with me different aspects of their work and some
of their worries. I hope I have not abused of their kindness.

I would like to gratefully acknowledge Professor William BOYES of the W. P. Carey

School of Business, Ms Melissa CLARK and Mrs Ana KUPRESANIN at ASU for
having contributed in different ways to the realization of this experimental work.

Beyond the thesis, these years of research where I benefited from the diversity and open
mindedness of the numerous people encountered in the field will remain an enriching
moment at a personal level. A moment I owe to my brother Tony COREY for having
familiarized me with the world of Wall St and for his constructive comments, but
primarily to my husband Nabeel for his unwavering confidence and constant support as
well as his patience especially during the difficult periods. I have particular thoughts for
my daughters Joyce and Martha for having read all or part of the thesis and for Ghassan,
Martha’s husband for his pertinent suggestions and remarks. I am profoundly grateful to
them. May they find herein the expression of all my love.

2
 A mon mari

3
REMERCIEMENTS
Je voudrais remercier en tout premier lieu le Professeur Mohammed ABDELLAOUI qui
a dirigé cette thèse pour son suivi et sa disponibilité pendant toute la durée de cette
recherche. Je lui sais gré de m’avoir fait bénéficier du logiciel qu’il a développé sans
lequel il aurait été difficile de mener à bien un travail expérimental dans le marché
boursier. Sa direction et sa confiance ainsi que ses explications claires et concises malgré
la distance géographique m’ont été particulièrement précieuces pour bien effectuer cette
recherche. Je remercie aussi vivement le Professeur Bertrand MUNIER pour m’avoir
acceuillie au GRID, pour sa gentillesse et son soutien tout le long de ce travail.

Ma reconnaissance va à ma famille aux Etats-Unis les KURI et à tous mes proches et

amis au Liban pour avoir personnellement contacté les gérants de portefeuilles pour les
entrevues. Sans pouvoir remercier nommément ces derniers, je voudrais leur exprimer
ma reconnaissance pour leur acceuil, leur coopération et leurs témoignages concernant les
différents aspects de leur travail. J’espère ne pas avoir abusé de leur gentillesse.

Je tiens également à témoigner ma reconnaissance au Professeur William BOYES à

l’Ecole W.P. Carey de Gestion, à l’Arizona State University ainsi qu’à Mademoiselle
Melissa CLARK et Madame Ana KUPRESANIN pour avoir contribué de façons
différentes à la réalisation de la partie expérimentale.

Au delà de la thèse, ces années de recherche où j’ai bénéficié d’une direction constructive
ainsi que de la diversité et de l’ouverture d’esprit des nombreuses personnes rencontrées
sur le terrain resteront un moment très enrichissant sur un plan personnel. Ce moment, je
le dois à mon frère Tony COREY pour m’avoir familiarisée avec le monde de Wall St et
pour ses commentaires constructifs mais surtout à la confiance et au soutien constants de
mon mari Nabeel ainsi qu’à sa patience aux périodes difficiles. Je sais gré à mes deux
filles Joyce et Martha d’avoir lu tout ou une partie de la thèse. J’ai également une pensée
particulière pour Ghassan le mari de Martha pour ses suggestions et ses remarques
pertinentes. Je leur suis profondément reconnaissante. Qu’ils trouvent ici l’expression de
toute mon affection.

4
TABLE DES MATIERES

INTRODUCTION GENERALE 63

PARTIE I: UTILITE ET PRISE DE DECISION 74

FACE AU RISQUE

CHAPITRE I. THEORIE DE L’UTILITE ESPEREE 79

1.1 HISTOIRE DU CONCEPT DE L’UTILITE 79

1.1.1 Origine de l’Utilité 80
1.1.2 Evolution 81
Bentham 81
La Révolution Marginaliste 81
La Révolution Ordinaliste 82
Le Concept d’Utilité dans les Recherches de vNM 83

1.2 UTILITE ESPEREE DANS LE RISQUE 85

1.2.1 L’Axiomatique de von Neumann et Morgenstern 85
L’Axiomatisation en termes de Mesures de Probabilité 85
Les Axiomes 86
Théorème Unique de l’Utilité Linéaire 88

1.3 UTILITE ESPEREE DANS L’INCERTAIN 89

1.3.1 L’Axiomatique de Savage 89
Les Axiomes 90
Le Thèorème de L’Utilité Espérée 91

5
1.3.2 L’Axiomatique de Anscombe et Aumann 93

CHAPITRE II. VIOLATIONS DE L’UTILITE ESPEREE ET LES 96

THEORIES ALTERNATIVES

2.1 VIOLATIONS DE CERTAINS AXIOMES DE L’UTILTE ESPEREE 96

2.1.1 Violation de l’Axiome d’Indépendence dans le Risque 97

Effet de la Conséquence Commune 97
Effet de Proportionalité 99
Transformation des Probabilités 99
2.1.2 Conséquence de la Violation pour le Choix Dynamique 102
2.1.3 Violation de l’Axiome d’Indépendence dans l’Incertain 105

2.2 VIOLATION DE LA STABILITE DES PREFERENCES 106

2.2.1 Transitivité 107
2.2.2 Invariance Procédurale 107
2.2.3 Invariance Descriptive 108
2.2.4 Le Point de Réflexion 108

2.3 LES THEORIES ALTERNATIVES 109

2.3.1 Sommaire des Développements Récents dans la Théorie de l’Utilité 109
2.3.2 Les Prédictions des Théories 111
Le Triangle des Probabilités 111
La Théorie de l’Utilite Espérée 112
La Théorie Originale des Prospects 113
L’Utilité dépendente des Rangs 116
La Théorie Duale 120
La Théorie (Cumulative) des Prospects 121

6
CHAPITRE III. LA THEORIE DES PROPSECTS 124

3.1 LE MODELE DESCRIPTIF 124

3.1.1 Les Eléments Non-Normatifs du Modèle 125

Le Point de Référence 125
La Localisation du Point de Référence 126
L’Aversion aux Pertes 126

3.1.2 La Fonction d’Utilité 128

3.1.3 La Fonction des Probabilités 129

3.2 LE MODELE FORMEL 131

3.2.1 Prise de Décision dans le Risque 131
3.2.2 Prise de Décision dans l’Incertain 133

3.3 L’AXIOMATIQUE DE LA THEORIE DES PROSPECTS 134

3.3.1 L’Idée de Tradeoff 134
3.3.2 L’Intuition 135
3.3.3 L’Axiomatization 136

3.4 LA CARACTERISATION DE L’ATTITUDE AU RISQUE 139

3.4.1 Les Notions d’Aversion au Risque 140
3.4.2 La Caractérisation de L’Aversion au Risque dans les Théories 143
La Caractérisation de L’Aversion au Risque sous EU 143
La Caractérisation de L’Aversion au Risque sous RDU 144
La Caractérisation de L’Aversion au Risque sous CPT 145

7
REMARQUES FINALES SUR LA PARTIE I 147

PARTIE II: ETUDE EXPERIMENTALE DES 150

FONCTIONS D’UTILITE ET DE L'AVERSION AUX PERTES

CHAPITRE IV. EXPERIENCE I: ELICITATION NON- 153

PARAMETRIQUE DE LA FONCTION D’UTILITE SUR LE TERRAIN

4.1 LE CADRE THEORIQUE 155

4.2 L’EXPERIENCE 156

4.2.1 La Procédure d’Elicitation 156
4.2.2 L’Application Expérimentale 159
Le Protocole Expérimental 159
Détails 161

4.3 ANALYSE DES DONNEES 165

4.3.1 Tests de Cohérence 165
4.3.2 Elicitation Non-Paramétrique des Fonctions d’Utilité 166
4.3.3 Ajustement Paramétrique des Données 167
4.3.4 Mesures de L’Aversion aux Pertes 169

4.4 RESULTATS CONCERNANT LA FONCTION D’UTILITE 172

4.4.1 Résultats des Tests de Cohérence 173
4.4.2 Tests de La Pondération des Probabilités 173
4.4.3 La Forme de la Fonction d’Utilité pour les Gains et les Pertes 173

8
4.5 RESULTATS CONCERNANT L’AVERSION AUX PERTES 177
4.5.1 Dans L’Agrégat 177
4.5.2 L’Aversion aux Pertes au Niveau de l’Individu 177

REMARQUES FINALES SUR L’EXPERIENCE I. 184

CHAPITRE V. EXPERIENCE II: EXPERIENCE DE 189

LABORATOIRE UTILISANT DES CONDITIONS DE
DOMINANCE STOCHASTIQUE

5.1 CRITERES DE DOMINANCE STOCHASTIQUE 191

5.1.1 Les Critères de Dominance Stochastique (L. et L., 2002) 191
L’Intuition 192
La Caractérisation des Préférences 193

5.1.2 Les Critères de Dominance Stochastique (B. et H., 2006) 196

L’Intuition 196
Incorporation de La Fonction de Transformation des Probabilités 197
Incorporation de L’Aversion aux Pertes 200

5.2 L’APPLICATION EXPERIMENTALE 202

5.2.1 L’Objectif 203

5.2.2 Les Sujets 203

5.3 ANALYSE DES DONNEES 204

9
5.4 RESULTATS CONCERNANT LA FONCTION D’UTILITE 206

5.5 RESULTATS CONCERNANT L’AVERSION AUX PERTES 207

5.6 RESULTATS CONCERNANT LA PROBABILITE GLOBALE DE 209

GAINS/PERTES

CONCLUSION 210

CONCLUSION GENERALE 212

COMMENTAIRES FINAUX ET PERSPECTIVES FUTURES 217

REFERENCES BIBLIOGRAPHIQUES 221

ANNEXE A: FIGURES 241

Figure 8: Equivalent Certain pour une Fonction d’Utilité Concave 241

ANNEXE B: ILLUSTRATION 242

ILLUSTRATION 2 Le Scroll Bar 242

ANNEXE C: TABLES 243

Table 5: Classification des Théories Alternatives 243
Table 8: Les Institutions Financières des Gérants de Portefeuilles 243
Table 10: Coefficients d’Aversion pour le Risque : Puissance et Exponentielle 244
Table 16(b) : sLes Paramètres des Gérants de Portefeuilles 245

10
ANNEXE D: LE QUESTIONNAIRE 247

LISTE DES FIGURES 254

LISTE DES TABLES 254

11
INTRODUCTION GENERALE

Cette recherche expérimentale a pour objectif principal l’élicitation des fonctions d’utilité
de gérants de portefeuille1 et la mesure de leurs degrés d’aversion aux pertes. Les
procédures d’élicitation présupposent que l’individu choisisse entre des alternatives
données comme s’il optimisait une fonction fondamentale de préférence. Généralement,
cette fonction de préférence résulte du modèle de l'utilité espérée (EU). Ce modèle
remonte à Cramer (1728) et Bernoulli (1738) qui ont proposé indépendamment que les
individus ne maximisent pas des valeurs monétaires espérées mais des utilités espérées,
c.-à-d. leur évaluation subjective des valeurs monétaires. Axiomatizé par von Neumann et
Morgenstern (1944) pour la prise de décision dans le risque et par Savage (1954) et
Anscombe et Aumann (1963) pour la prise de décision dans l'incertain, EU est le modèle
normatif de choix qui décrit la manière selon laquelle les individus doivent prendre leurs
décisions. Cependant pour qu’un modèle soit opérationnel et prescriptivement utile en
analyse de décision, le comportement actuel des individus dans des situations de choix
simples doit être compatible avec le comportement supposé dans le modèle. Néanmoins,
EU s’est avérée indéfendable comme modèle descriptif du comportement des décideurs.
En effet, les travaux expérimentaux qui ont suivi ces travaux théoriques ont montré que
les gens violent systématiquement certains axiomes de EU (e.g. Allais, 1953; Ellsberg,
1961; Kahneman et Tversky, 1979).

L'évidence empirique a motivé les chercheurs à développer des théories alternatives de

choix qui tiennent compte des violations observées. Ces modèles appelés “utilité non-
espérée” ou “généralisations de l'utilité espérée” ont été examinés plus tard dans le
triangle unité pour trouver si possible le modèle qui se conforme le plus au
comportement réel (Harless et Camerer, 1994; Hey et Orme, 1994; Carbone et Hey,

1
Gérant de portefeuille est pris ici dans le sens général de responsable de la diversifaction d’un portefeuille.
Il renvoie aux conseillers financiers et aux gérants de fonds.

12
1995). Le triangle unité est l'ensemble des distributions des probabilités ou loteries qu'on
peut définir sur un ensemble de trois résultats différents. Ce simplexe est généralement
représenté par le triangle unité dans un système de deux axes rectangulaires. Si les
résultats sont fixes, le triangle contient l'ensemble de toutes les loteries possibles avec le
bord gauche caractérisé par une probabilité zéro pour la meilleure conséquence, le bord
inférieur par une probabilité zéro pour la moins préférée, et l’hypoténuse par une
probabilité zéro pour la conséquence moyenne. Néanmoins, aucune des généralisations
de l'utilité espérée ne s'est avérée sensiblement plus conforme que EU aux données
empiriques dans le triangle entier, c.-à-d. dans toutes les situations risquées possibles.
Selon Abdellaoui et Munier (1998), la raison en est que les préférences des décideurs
dépendent du risque des prospects auxquels ils font face, c.-à-d. de la région du triangle
unité qui représente leur situation. Un des résultats de leur expérience est que les modèles
d’utilité non-espérée décrivent les préférences des individus mieux que EU en dehors du
milieu du triangle. L’enquête de Starmer (2000) qui évalue les données des expériences
dans le triangle unité présente le modèle d'espérance d’utilité à dépendance du rang qui
ne viole pas la monotonicité comme le "meilleur pari" parmi les alternatives en dehors de
l'intérieur de la triangle. Dans ce modèle les probabilités sont remplacées par des
pondérations qui reflètent l'impact des événements sur la désirabilité des conséquences.
Ces poids de décision qui résultent de la pondération de probabilités cumulées et dont la
somme est égale à un sont assignés aux conséquences selon leurs rangs respectifs dans les
séries.

Parmi les modèles à dépendance du rang, la théorie des prospects (CPT)2 de Tversky et
Kahneman (1992) c.-à-d. la version révisée de leur théorie originale (1979), explique
également les régularités du comportement sur le terrain considérées des paradoxes sous
EU. Camerer (2000) montre que dans plusieurs domaines CPT explique les anomalies et
les phénomènes de base que EU est utilisée pour expliquer grâce aux éléments
psychologiques qui y sont intégrés.

2
CPT pour cumulative prospect theory.

13
LES RÉSULTATS EXPÉRIMENTAUX

Esquissées en bref, les caractéristiques principales de la théorie des prospects sont: 1) la

fonction d’utilité définie sur des gains et des pertes relatifs à un point de référence neutre
exhibe une sensibilité marginale décroissante aux déviations du point de référence
donnant lieu à une fonction d’utilité en forme de S c’est-à-dire concave pour les gains
(u’’ < 0) et convexe pour les pertes (u’’> 0); 2) la fonction d’utilité est plus pentue dans
le domaine des pertes que dans le domaine des gains quand la valeur d'une perte est plus
faible en valeur absolue que la valeur d'un gain commensuré, une caractéristique marquée
par les auteurs comme aversion aux pertes; et 3) les décisions sont basées sur des
distributions cumulatives subjectives données par l’équation: π = w( p ) où p renvoie à la
distribution cumulative objective, w à un traitement subjectif de p tels que w’(.) > 0, w (0)
= 0 et w (1) = 1. Les éléments psychologiques intégrés dans la théorie sont par
conséquent: l'effet de réflexion, l’aversion aux pertes et le traitement subjectif des
probabilités. (La théorie des prospects est élaborée au chapitre III, p: 124 du document en
anglais).

L'effet de réflexion est un élément psychologique central de la théorie des prospects.

D’après Kahneman et Tversky (1979, p: 268), lorsqu’on passe d’un contexte de gains à
un contexte de pertes, il y a renversement de préférences. En d’autres termes, si les
individus préfèrent des gains sûrs à des gains probables, ils préféreraient des pertes
probables aux pertes sûres. L'effet de réflexion explique l'hésitation quant à la réalisation
des pertes au marché boursier: Terence Odean (1988) constate que comme Shefrin et
Statman (1985) ont prévu, les investisseurs réalisent leurs gains trop tôt et sont peu
disposés à réaliser leurs pertes excepté pour le mois de décembre (et ce pour des raisons
d’impôts) et que leur comportement les a menés à de moindres bénéfices.
Graphiquement, les investisseurs auraient une fonction d’utilité concave pour des gains et
convexe pour des pertes. Les élicitations basées sur la méthode du tradeoff qui filtre la
pondération des probabilités (Wakker et Deneffe, 1996) généralement corroborent la
concavité pour les gains dans l’agrégat et pour la plupart des individus (Abdellaoui, 2000;
Abdellaoui, 2002; Bleichrodt et Pinto, 2000). L'évidence pour la convexité pour des

14
pertes est cependant, moins tranchante au niveau de l’individu (Abdellaoui, 2000;
Fenema et van Assen, 1999) et d'ailleurs, Levy et Levy (2002) trouvent une utilité
concave pour des pertes et une utilité convexe pour des gains pareille à celle de
Markowitz (1952).

L’autre élément central des résultats psychologiques de la théorie des prospects est
l’aversion aux pertes. Cet élément renvoie à l'assymétrie des gains et des pertes
relativement à un point de référence avec les pertes apparaissant plus grandes que les
gains commensurés. Kahneman et Tversky (1979, p: 279) définissent l'aversion aux
pertes comme suit: un individu est averse aux pertes s'il n’aime pas les paris symétriques
et si en plus, l'aversion pour de tels paris augmente avec la valeur absolue des enjeux. Les
auteurs montrent aussi que leur définition est équivalente sous l’hypothèse de la théorie
des prospects à une fonction d’utilité plus pentue pour les pertes que pour les gains. Le
dernier résultat est à la base des diverses définitions de l'aversion aux pertes qui ont suivi:
Tversky et Kahneman (1992); Wakker et Tversky (1993); Bowman et al. (1999);
Köbberling et Wakker (2005); et Neilson (2002). Ces définitions seront présentées et
opérationalisées ultérieurement (p: 34).

L’appui théorique pour l'aversion aux pertes vient de Rabin (2000) qui prouve que des
degrés plausibles d'aversion au risque pour des enjeux modestes sous EU impliquent des
degrés élevés irréalistes d'aversion au risque pour de grands enjeux tandis que l'aversion
aux pertes implique une aversion au risque pour des enjeux modestes et une plausible
aversion au risque pour de grands enjeux. La suggestion de Rabin favorisant l'utilisation
de l'aversion aux pertes pour expliquer l'aversion au risque a été réitérée dans
"Anomalies: Aversion au Risque" par Rabin et Thaler (2001).

L’appui empirique pour l'aversion aux pertes vient de Putler (1992) et de Hardie,
Johnson, et Fader (1993) pour les élasticités assymétriques des prix. L'aversion aux pertes
explique aussi l'effet de dotation (Thaler, 1980; Loewenstein et Adler, 1995) et par
conséquent la disparité entre la bonne volonté de payer et la bonne volonté d'accepter
(Kahneman, Knetsch et Thaler, 1990; Bateman, Munro, Rhodes, Starmer et Sugden,

15
1997), le biais du status quo (Samuelson et Zeckhauser, 1988), et l'effet de disposition
(Weber et Camerer, 1988, Odean, 1998; Heath, Huddart et Lang, 1999). Selon Benartzi et
Thaler (1997) et Barberis, Huang et Santos (2001) l'aversion aux pertes est nécessaire
pour expliquer l’equity premium puzzle, c’est à dire les primes que les investisseurs
exigent pour investir en actions et qui sous l'utilité espérée impliquent un degré d'aversion
au risque absurde ainsi qu’a été démontré par Mehra et Prescott (1985).

Le troisième résultat expérimental intégré dans la théorie est la pondération non-linéaire

des probabilités. Contrairement à EU où l'utilité (de chaque conséquence possible) est
pondérée par sa probabilité, dans la théorie des prospects, l'utilité est multipliée par un
poids de décision π (p) qui est une fonction strictement croissante de p mais qui n'est pas
une probabilité (Tversky et Kahneman, 1986). Les poids de décision ont été introduits
pour modeler la tendance à surpondérer les faibles probabilités et souspondérer les
probabilités modérées et grandes. Preston et Baratta (1948) ont été les premiers à
observer cette tendance. Plus récemment, Diamond (1988) et Edwards (1996) ont
constaté que des sujets jugeant des risques composés de faible probabilité/grande
conséquence ont été plus affectés par la grandeur de la conséquence que par la probabilité
tandis que ceux jugeant des risques composés de grande probabilité/faible conséquence
ont eu tendance à combiner les deux composantes. Selon MacCrimmon et Larsson (1979)
les sujets ont tendance à choisir la loterie avec le gain plus probable pour des probabilités
élevées de gain et à choisir la loterie avec le gain le plus élevé pour de faibles probabilités
de gain.

D’autres appuis empiriques ont été donnés plus tard par des études paramétriques qui
supposent une fonction de probabilité pondérée spécifique: Tversky et Kahneman (1992);
Camerer et Ho (1994); Tversky et Fox (1995); Wu et Gonzalez (1996); Abdellaoui
(2000) et par des études non-paramétriques: Abdellaoui (2000) et Bleichrodt et Pinto
(2000).

16
MOTIVATION ET OBJECTIFS DE RECHERCHE

La popularité de CPT et la possibilité d'expliquer les anomalies sous EU par des

préférences averses aux pertes ont inspiré des économistes à rechercher des méthodes qui
peuvent éliciter des fonctions d’utilité sous CPT, tester l'aversion aux pertes et mesurer
son degré. Parmi les études qui ont examiné l’aversion aux pertes au niveau individuel on
peut distinguer: Bleichrodt et Pinto (2002); Schmidt et Traub (2002) qui ont réalisé plutôt
des tests qualitatifs et Abdellaoui et al. (2006) qui ont mesuré quantitativement l’aversion
aux pertes.

La disponibilité de la méthode non-paramétrique d'Abdellaoui et al. (2006) et la facilité

de son applicabilité (une forme automatisée avec le graphe de l'utilité élicitée
simultanément pour des gains et des pertes obtenu immédiatement à la fin de
l’expérience) dans un temps relativement court a offert la possibilité d’éliciter les
fonctions d'utilité de gérants de portefeuilles pour qui le temps est une denrée rare. Les
entrevues étant conduites dans les bureaux des gérants de portefeuilles à leurs institutions
financières respectives, les élicitations ont exigé beaucoup de déplacements; néanmoins,
les avantages potentiels des élicitations importent aux économistes intéressés par la
modélisation du comportement, aux chercheurs des anomalies sur le marché boursier et
aux analystes de décision particulièrement s'ils sont corroborés par d'autres investigations
sur le terrain.

En effet, l’élicitation de la forme de la fonction d’utilité fournit une perspective

descriptive aux modélisateurs d'une règle rationnelle pour la prise de décision, étant
donné que les règles normatives doivent tenir compte du comportement actuel de
l’individu (Allais, 1953; 1979).

En plus, étant non-paramétrique et donc indépendante de tout choix de fonctionnelle,

l’élicitation de la fonction d’utilité serait aussi utile pour les chercheurs intéressés par les
anomalies du marché financier. Par exemple, Benartzi et Thaler (1997) ont employé les

17
évaluations de Tversky et Kahneman (1992) pour les coefficients d'aversion aux pertes et
les fonctions d’utilité d'investisseurs afin de calculer les valeurs espérées des prospects,
actions et obligations visant à expliquer le paradoxe des primes pour les actions. La
méthode d’élicitation de Tversky et Kahneman (1992) suppose une forme paramétrique
“puissance” pour la fonction d'utilité, l'évaluation de tous les paramètres étant
problématique à ce moment-là (ibid, p: 311).

Enfin, un autre avantage est dans l'analyse de décision où les utilités biaisées peuvent
avoir comme conséquence des prévisions économiques inexactes. Traditionnellement,
l'analyse de décision suppose la normative EU pour l'élicitation des fonctions d’utilité.
Cependant, pour que le postulat soit valide prescriptivement les préférences du décideur
doivent être compatibles avec EU. Une utilisation corrective de la théorie des prospects a
été suggérée par Kahneman et Tversky (1979); von Winterfeld et Edwards (1986);
Fischnoff (1991); et Kahneman et Tversky (2000, p: 157). Bleichrodt, Pinto et Wakker
(2001) proposent l'utilisation des utilités corrigées dans les prescriptions des décisions
optimales, la correction basée sur des paramètres de préférence trouvés empiriquement.

ANNONCE DU PLAN

Ce travail expérimental élicite donc les fonctions d’utilité des gérants de portefeuilles et
mesure leurs degrés d’aversion aux pertes sous CPT. Les choix présentés à ces derniers
ayant été construits dans le cadre de la théorie de la prise de décision dans le risque, ce
travail est divisé en deux parties: La première partie présente le cadre théorique de la
prise de décision dans le risque et construit la scène pour le travail expérimental de la
deuxième partie.

La première partie se compose d’une introduction et de trois chapitres. L’introduction

distingue les étapes principales dans l’évolution de l’utilité espérée jusqu'au travail
pionnier du vNM (1944). Le premier chapitre présente l’axiomatization de l'utilité
espérée par ces derniers auteurs pour le risque et celle par Savage (1954) pour l'incertain.
Le deuxième chapitre explore les violations de certains axiomes de EU, présente les

18
théories alternatives de la prise de décision et montre que la théorie des prospects (1992)
explique ces violations et l’évidence empirique de phénomènes tels que l’aversion aux
pertes et l’effet de réflexion par l’intégration d’éléments non-normatifs. Le troisième
chapitre explore ces éléments avant de présenter le modèle formel, son axiomatization
dans le risque ainsi que la caractérisation de l'aversion au risque sous les différentes
théories.

La deuxième partie présente le travail expérimental. D’abord, les fonctions d’utilité des
gérants de portefeuilles sont élicitées et leurs coefficients d'aversion aux pertes mesurés
suivant la méthode d'Abdellaoui et al. (2006). Les résultats sont ensuite contrastés avec
ceux de ces derniers pour étudiants. Ensuite, la méthode de Baucells et Heukamp (2006)
est employée pour examiner les préférences d’étudiants en MBA. A chaque expérience,
un chapitre est consacré qui commence par une brève introduction montrant le
dévelopment progressif de la méthode de celles qui l’ont précédées. L’introduction est
suivie de la description de la procédure d’élicitation et de l’application expérimentale.
Les résultats sont ensuite contrastés avec ceux de la littérature récente. La conclusion
générale est suivie de commentaires finaux et perspectives futures.

19
PARTIE I: THEORIE DE L’UTILTE ET PRISE DE DECISION FACE AU
RISQUE

Cette partie présente d’abord EU qui est un point de départ normal puisque les théories
alternatives qui suivent sont des généralisations de cette théorie de base. La présentation
de ces alternatives est cependant restreinte à deux égards: 1) le risque est distingué de
l'incertain et la concentration est sur la décision dans le risque. En effet, l'objectif de la
première partie est de présenter le cadre théorique qui convient aux essais empiriques
présentés dans la deuxième; or, les choix présentés aux gérants de portefeuilles et aux
étudiants en MBA pour inférer leurs préférences ont été conçus pour des situations de
risque; 2) l'examen des alternatives se concentre sur des modèles présupposant une
fonction simple de préférence, mais défendables en tant que modèles descriptifs du
comportement réel. Ce postulat est un principe important de cohérence et il est
raisonnable de supposer que les gens souhaitent y obéir même s’il est exigeant. En outre,
bien que les violations empiriques d'une fonction simple de préférence présentent un cas
pour les modèles non-conventionnels quand ceux-ci ne sont pas jugés en utilisant des
critères raisonnables, abandonner la notion de préférences bien définies exige des
changements qui augmentent la complexité de la théorie, réduisent son rendement
prédictif et la rendent moins compatible avec le reste de la théorie économique (Starmer,
2000). Selon Arrow (1995) “ces modèles sont susceptibles d'être très corrects, c’est juste
que leurs prévisions sont beaucoup plus vagues que celles suggérées par la rationalité; la
rationalité est unique.” Ainsi restreinte, cette partie se compose de trois chapitres: le
premier est consacré à EU et comprend l’histoire du concept d’utilité et le modèle formel;
le deuxième est consacré aux violations de certains axiomes de ce modèle et aux
alternatives dévelopées en réponse et le troisième à la théorie de prospects (1992) qui tout
en étant cohérente est la plus valide descriptivement grâce aux éléments non-normatifs
intégrés dans la théorie; ce dernier chapitre de la première partie comprend le modèle
descriptif, le modèle formel ainsi que la caractérisation de l’aversion au risque sous les
différentes théories.

20
CHAPITRE I. LA THEORIE DE L’UTILITE ESPEREE

Ce chapitre commence par situer historiquement le concept d’utilité et établit un langage

commun pour le reste du document. Il distingue les étapes principales dans l’évolution
de la EU jusqu'au travail pionnier du vNM (1944) qui en constitue la version moderne.
Ainsi, il montre comment le concept débute avec Bernoulli (1738), évolue avec Bentham
(1789) et surtout avec la révolution des marginalistes au dix-neuvième siècle période à
laquelle la définition de l’utilité marginale fut établie. Celle-ci est suivie bientôt par la
révolution des ordinalistes au vingtième. L’utilité dans le certain et les comparaisons
cardinales sont pratiquement abandonnées en faveur d’une vision ordinaliste de l’utilité
où le principe de l’utilité marginale décroissante est cependant implicite. Finalement, le
concept moderne de vNM (1944) est introduit avant de présenter le modèle formel.
(L’histoire du concept de l’utilité est traitée dans la section 1.1 du document en anglais,
p: 79).

Plusieurs axiomatisations dans le risque ayant été proposées (e.g. Herstein et Milnor,
1953; Jensen, 1967, Luce et Raiffa, 1957), l'axiomatisation présentée est basée sur
Fishburn (1970). D'abord, le cadre général et quelques définitions essentielles sont
donnés; les axiomes sont ensuite exposés de manière formelle et le théorème de
représentation énoncé. (L’axiomatization est présentée dans la section 1.2, p: 85).

La représentation plus générale de l’utilité espérée subjective (SEU) de Savage (1954)

suit. Elle peut être considérée comme une combinaison de la théorie de vNM (1944) et de
sa duale la théorie de probabilité subjective de Bruno de Finetti (1937). Par rapport à la
construction de vNM (1944), il ya donc plus de conditions indiquées. Savage y énonce
les conditions qui permettent de montrer l’existence d’une mesure de probabilité
subjective sur l’ensemble des états de la nature (un état de la nature étant une description
complète d’une situation possible de l’environnement du décideur) et d’une fonction
d’utilité similaire à celle de vNM sur l’ensemble des conséquences communes.

21
L'axiomatisation basée sur Fishburn (1970) est présentée dans la section 1.3. Anscombe
et Aumann (1963) ayant suivi une route intermédiaire entre vNM et Savage, une courte
description de leur théorie est esquissée.

Les axiomes présentés dans ce chapitre ont une énorme attraction normative et EU
semble pouvoir être utilisée sans beaucoup de difficultés. Ces travaux théoriques ont été
cependant suivis par des travaux expérimentaux qui ont montré que certains des axiomes
de cette théorie sont violés systématiquement. Le deuxième chapitre développe en détail
ces violations et présente les théories dévelopées comme alternatives.

CHAPITRE II. VIOLATIONS DE L’UTILITE ESPEREE ET THEORIES

ALTERNATIVES

Ainsi que le titre l’indique, ce chapitre est divisé en deux parties: la première est
concernée par les violations du modèle de l’utilité espérée face au risque et la deuxième
par les théories dévelopées en alternatives.

LES VIOLATIONS
C’est la violation de l'axiome de l'indépendance dans le risque qui est la plus discutée en
littérature et qui est également responsable de la génération de beaucoup d’alternatives à
EU et à SEU pendant une longue période s'étendant de 1979 jusqu'à ce jour. C’est que
comme l’explique Fishburn (1970) le principe d'espérance et la linéarité dans les
probabilités ne peuvent pas être retenus sans cet axiome. Ce chapitre présente d’abord les
violations de cet axiome face au risque (Allais, 1953) pour expliquer ensuite comment la
théorie des prospects (1992) tient compte de ces violations. Une brève description de la
violation de cet axiome en contexte dynamique et dans l’incertain suit.

22
La première section commence par décrire les deux exemples conçus par Allais (1953)
qui le premier démontre que la propriété de séparabilité3 que l'axiome de l'indépendance
implique est violée. En effet, ces deux exemples montrent la violation de cette propriété
qui est additive et multiplicative et sont connus sous le nom d’effet de la conséquence
commune et effet de proportionalité respectivement. Ces derniers sont présentés dans la
section 2.1.1, Tables 1 et 2 respectivement, p: 97, 99 respectivement). Cette première
section montre aussi pourquoi les deux effets violent l’axiome d’indépendance et
pourquoi les préférences sont incohérentes sous EU alors que la section suivante explique
comment la théorie des prospects (1992) tient compte de ces violations (p: 99).

En effet cette dernière incorpore une pondération non-linéaire des probabilités, p → w(p)
également appelée une fonction de transformation des probabilités. Deux propriétés
exigées sur cette fonction réconcilient les préférences d’Allais qui sont contradictoires
sous EU: La sous-additivité explique la violation de la conséquence commune et la sous-
proportionalité explique l’effet de proportionalité (Prelec, 2000). La fonction de
transformation des probabilités a une propriété empirique en plus et “peut-être la plus
importante” (Prelec, 2001) qui indique que les petites probabilités sont surpondérées et
les grandes probabilités sont souspondérées.

La section suivante montre ce que la violation de l’axiome d’indépendance implique pour

le choix dynamique. Connue dans ce contexte comme violation de conséquentialisme,
elle implique qu'au moins un principe de choix dynamique est violé puisque dans ce
contexte quatre conditions impliquent conjointement l'équivalence de l'indépendance
(cette section basée sur Wakker (1999) est présentée aux pages 103- 105 avec des figures
pour illustrer).

Finalement, les violations des axiomes responsables de la stabilité des préférences sont
présentées (p:107) c.-à-d. la transitivité, l'invariance descriptive et l'invariance de
procédure. Bien que pour Ramsey (1931) la possibilité que le choix dépende de la forme

3
Cette propriété est justifiée par l'exclusivité mutuelle des conséquences de la loterie (Machina, 1989;
Weber et Camerer, 1987).

23
spéciale des options offertes soit “absurde”, l’exemple de Kahneman et de Tversky
(1979) connu sous le nom de la maladie d’Asie prouve que les gens sont influencés par le
contexte: les choix sont renversés quand des conséquences initialement présentées en
termes de vies sauvées sont présentées en termes de vies perdues. L'inversion de
préférence implique que le postulat de l'invariance descriptive est violé. L'effet de
réflexion c.-à-d., le renversement des préférences quand on passe d’un contexte de gains
à un contexte de pertes est un exemple compatible avec l’effet de contexte.

En raison des violations ci-dessus, beaucoup d'économistes ont conclu que le modèle de
l’utilité espérée ne correspond pas du tout aux faits ou ne correspond pas seulement à
certains de ces faits, en dépit des arguments qui procèdent avec une logique sans faute
des postulats à la conclusion. Pour le développement d'un concurrent sérieux au moins
pour quelques objectifs, les théoriciens de décision ont revisité et révisé les axiomes de
l'EU et ce faisant ont généré un grand nombre de théories allant de pair avec les essais
expérimentaux continus de ces théories.

LES THEORIES ALTERNATIVES

Basée sur les revues de ces théories alternatives et des essais expérimentaux les
concernant (Camerer, 1989; Schmidt, 2002; et Starmer, 2000) cette partie du chapitre se
concentre sur les modèles qui selon ces revues expliquent le mieux les données
empiriques actuellement disponibles. Ainsi les théories alternatives présentées dans cette
section sont limitées aux modèles à dépendance du rang (Quiggin, 1982; Tversky et
Kahneman, 1992) qui ne sont pas linéaires dans la probabilité, particulièrement en raison
du degré saisissant de convergence à travers les études concernant la forme de leurs
fonctions de transformation des probabilités (Starmer, 2000, p: 359).

Le souci principal étant la réconciliation des prévisions de ces théories avec les faits
expérimentaux, la présentation est restreinte à une comparaison dans le triangle unité des
formes des courbes d'indifférence entre les ensembles de loteries sous les différentes
théories et celles conjecturées des observations au laboratoire et/ou sur le terrain. Les
courbes d'indifférence sous EU sont discutées en premier lieu (à la page 113 du

24
document) pour être ensuite contrastées avec celles générées des choix qui violent
l’axiome d'indépendance. Les courbes d'indifférence sous RDU (Quiggin, 1982) et CPT
(Tversky and Kahneman, 1992) suivent (p: 116, 121 respectivement).

Pour souligner cependant, les inspirations et les idées que les auteurs ont tirées l'un de
l'autre et les sauts intuitifs dans la pensée convergente favorisés par la combinaison de la
connaissance existente et/ou la tolérance de la dualité, la théorie des prospects (1979), la
version originale de la théorie des prospects (1992) est présentée d'abord (p: 113). La
théorie anticipée de Quiggin (1982), développée en partie pour inclure certaines
caractéristiques de la théorie non-conventionnelle des prospects tout en suivant une
stratégie conventionnelle est présentée ensuite. La théorie duale de Yaari (1987), un cas
spécial de celle de Quiggin, qui est linéaire dans les conséquences et non-linéaire dans les
probabilités et donc la duale de EU, est décrite brièvement (p: 120) suivie finalement par
la théorie des prospects (1992). En résumé, cette première partie montre que la théorie
des prospects réussit à expliquer les anomalies et les phénomènes de base que EU est
utilisée pour expliquer grâce à une combinaison de réalisme empirique et d’avantages
théoriques qui favorisent son utilisation comme approximation de la fonction de
préférence que l’individu est supposé optimiser. En raison de la pertinence des éléments
psychologiques de cette théorie à la partie expérimentale le troisième chapitre présente le
modèle descriptif ainsi que le modèle formel et son axiomatisation pour le risque.

CHAPITRE III . LA THEORIE DES PROSPECTS

La théorie des prospects (1992) en tant qu’une fertilisation croisée de la théorie des
prospects (1979) et de celle de Quiggin (1982) présente un mix de validité descriptive et
de précision mathématique. Ce chapitre souligne en premier lieu les éléments non-
normatifs de la théorie avant de présenter le modèle formel et son axiomatisation dans le
risque. Ces éléments psychologiques résultent de ce que l’individu perçoit les
conséquences/probabilités relativement à un point de référence. L'intuition de ce point
est donnée d’abord et l'importance de sa localisation pour l'ordre des préférences est
soulignée suivie d’une section consacrée à l'aversion aux pertes (p: 125, 126

25
respectivement). Les implications de ces éléments non-normatifs pour la fonction
d’utilité sont détaillées ensuite (p: 128).

Bien que ce travail soit concerné principalement par l'élicitation de la fonction d’utilité, la
fonction de transformation des probabilités en forme de “S inversé” et différenciée pour
les gains et les pertes est décrite (p: 129) afin de souligner l'importance de filtrer son
impact sur la désirabilité des conséquences.

Le modèle formel pour le risque ainsi que celui pour l'incertain sont présentés suivis de
l'axiomatisation de la théorie des prospects (1992) dans le risque basée sur Châteauneuf
et Wakker (1999). L’idée centrale de l'axiomatisation de CPT dans l’incertain et dans le
risque est la tradeoff consistency ou cohérence du tradeoff. Cette dernière est également
essentielle à la méthode d'élicitation employée dans ce travail expérimental pour encoder
les fonctions d’utilité des gérants de portefeuilles. Par conséquent, la définition de l'idée
du tradeoff basée sur Wakker (1994) est d'abord présentée (p: 134). L'intuition pour la
cohérence du tradeoff est ensuite illustrée par un exemple (p: 135) pris de Wakker et de
Tversky (1993, p: 149). L'axiomatisation pour le risque suit (p: 136).

Ce chapitre présente également les caractérisations respectives de l'aversion pour le

risque sous EU, RDU, et CPT pour montrer que ce qu’on comprend par aversion au
risque a été largement raffiné dans le cadre de RDU et d’avantage dans celui de CPT.

Les notions de l'aversion de risque définies indépendamment de n'importe quel modèle

sont présentées dans une première étape (p: 140) et les liens entre ces notions et les
fonctions de préférence sous les différentes théories sont établis dans une deuxième (143-
145). Ainsi, des comportements de risque observés de nature différente qui ne sont pas
distingués sous EU sont séparés sous RDU et encore plus sous CPT où l’aversion au
risque a trois composantes: une fonction d’utilité, une fonction de transformation des
probabilités et une aversion aux pertes.

26
En somme, en plus des avantages théoriques qui caractérisent les théories dépendantes du
rang, le réalisme empirique de la CPT la prédispose à être une bonne approximation de la
fonction fondamentale de préférence que l'investisseur est supposé optimiser. Comme
c’est aux mesures des paramètres de la théorie de montrer si l'approximation est bonne, la
première expérience élicite la fonction d’utilité des gérants de portefeuilles et mesure
leurs coefficients d'aversion aux pertes sous CPT testant ainsi la théorie sur le terrain. Les
résultats sont ensuite contrastés avec ceux d’Abdellaoui et al. (2006) pour étudiants. Une
deuxième expérience examine qualitativement les préférences d’étudiants en MBA
également sous CPT (Baucells et Heukamp, 2006). Combinées, les deux expériences
offrent une comparaison entre gérants de portefeuille expérimentés et en potentiel.

27
PARTIE II: ETUDE EXPERIMENTALE DES FONCTIONS D’UTILITE
INDIVIDUELLES ET DE L'AVERSION AUX PERTES

La plupart des méthodes d'élicitation employées dans les études empiriques qui
s’intéressent à la fonction fondamentale de préférence de l’individu ont supposé des
formes paramétriques spécifiques pour la fonction d’utilité rendant de ce fait les
inférences au sujet de ces fonctions dépendantes du choix des fonctionelles. Aussi, la
méthode choisie pour être utilisée dans cette la première expérience partie est-elle non-
paramétrique. En plus, la pondération de la probabilité ne lui pose pas de problème. La
pondération de la probabilité est une cause importante des violations de la théorie de
l’utilité espérée et des contradictions systématiques parmi les différentes méthodes
d'élicitation d’utilité qui devraient donner le même résultat sous EU (Hershey et
Schoemaker (1985); McCord et de Neufville (1986); Wakker et Deneffe (1996);
Bleichrodt et al. (2001); Fischoff (1982); Schkade (1988)). La robustesse contre la
pondération de la probabilité est fondamentale pour que la méthode reste valide pour
RDU et CPT et puisse être appliquée dans l'analyse prescriptive de décision.

Aussi, la méthode d'Abdellaoui et al. (2006) est-elle employée pour éliciter entièrement
les fonctions d’utilité des gérants de portefeuilles permettant ainsi la mesure de leurs
coefficients d'aversion aux pertes. Quand aux préférences des étudiants en MBA, elles
seront inférées à partir de loteries construites par Baucells et Heukamp (2006). A chaque
expérience un chapitre est consacré qui commence par une brève introduction qui montre
que chaque méthode contient tous les éléments essentiels de celles qui l’ont précédée
mais non vice versa. Autrement dit, c’est le cas où tout ce qui est inférieur est contenu
dans ce que est supérieur mais tout ce qui est supérieur n’est pas contenu dans ce qui est
inférieur comme Aristote a été le premier à le préciser. A cet égard, pour déterminer ce
qui est inférieur dans une séquence, Wilber (1996) suggère de réfléchir à un cas où tout
ce qui est supérieur est détruit mais rien de qui est inférieur ne l’est. L’introduction est
suivie de l’application expérimentale de la méthode pour chaque chapitre.

28
CHAPITRE IV. EXPERIENCE I: ELICITATION NON-PARAMETRIQUE DE
LA FONCTION D’UTILITE SUR LE TERRAIN

Dans ce chapitre, les fonctions d’utilité des gérants de portefeuilles sont élicitées
simultanément pour les gains et pour les pertes et leurs coefficients d'aversion aux pertes
mesurés en utilisant la procédure d'élicitation d'Abdellaoui et al. (2006). La première
section récapitule la théorie des prospects; la deuxième décrit la procédure d'élicitation
(p: 156) ainsi que l’application expérimentale (p: 159); quand à la troisième, elle est
concernée par l'analyse des données et comprend les élicitations non-paramétriques,
l'ajustement paramétrique et la mesure de l'aversion aux pertes dans l'agrégat aussi bien
qu’au niveau individual (p: 165-172). Les sections suivantes présentent les résultats tout
en les contrastant avec les résultats d’études précédentes; une section finale discute les
résultats de l’expérience (p:184).

LA PROCEDURE D’ELICITATION

La méthode est basée sur l’élicitation de points d’utilité, mipoints pour être précis qui est
souvent utilisée dans l’axiomatisation des modèles de décision. La procédure consiste en
quatre étapes et elle est résumée dans la Table 1 qui suit. La deuxième colonne décrit la
quantité élicitée, la troisième l’indifférence qu’on recherche et la quatrième ce que cette
indifférence implique sous l’hypothèse CPT. La dernière colonne montre les variables qui
doivent être spécifiées et les valeurs choisies pour ces variables.

29
 Table 1: Procédure d’Elicitation
Quantité
Élicitée Indifférence Sous l’Hypothèse CPT Variables de Choix
Etape 1 L1 (L1, p’; L*) ∼ (L0, p’; L) p’ = 0.33
L2 (L2, p’; L*) ∼ (L1, p’; L) U(L0) - U(L1) = U(L1) -U(L2) L* =-100
pl L1 ∼ (L2, pl ; L0) w-(pl) = 0.5 L=-600, L0 = -1000
G1 (G1, p’; G) ∼ (G0, p’; G*) p’ = 0.33
G2 (G0, p’; G) ∼ (G1, p’; G*) U(G2) - U(G1) = U(G1) - U(G0) G* = 600
+
pg G1 ∼ (G2, pg; G0) w (pg) = 0.5 G = 100, G0 = 1000
Etape 2 Lr ∈ [ L1 , 0] Lr ∼ (LA, pl ; LB) U(Lr) = 0.5 U(LA) + 0.5 U(LB) L1 = -100000
-
Etape 3 l Ls ∼ (l, 0.5 ; 0) w (0.5)U(l) = -s s = 0.25
g 0 ∼ (g, 0.5 ; l) w+(0.5)U(g) = s
Gs Gs∼ (g, 0.5 ; 0) U(Gs) = w+(0.5)U(g) = s
Etape 4 Gr ∈ [ 0, Gs] Gr ∼ (GA, pg ; GB) U(Gr) =0.5 U(GA) + 0.5 U(GB)

Les Etapes
D’abord, deux conséquences monétaires pour les gains G1, G2 et deux pour les pertes L1,
L2 pour lesquelles la différence d’utilité entre les conséquences successives est constante,
sont déterminées en séquence. Ensuite deux probabilités pg et pl pour lesquels w(pg) =
w(pl ) = 0,5 sont déterminées par la méthode de l’équivalence des probabilités. L'étape
suivante, fractile pertes détermine une séquence d'onze conséquences pour les pertes Lr
élicités sur l’intervalle [-100000, 0] pour les utilités suivantes: 0.015, 0.031, 0.062,
0.093, 0.125, 0.25, 0.375, 0.5, 0.625, 0.75, 0.875, où U (Lr) = - r en utilisant la probabilité
élicitée pour les pertes.

La troisième étape détermine le gain x8 qui représente la même variation d’utilité par
rapport à la conséquence 0 comme valeur absolue de la variation d’utilité entre la perte
L0.25 et 0; cette étape lie l'utilité pour les pertes à l'utilité pour les gains à travers
l'élicitation de trois conséquences et permet la mesure de l'aversion aux pertes.

La quatrième étape, appelée fractile gains élicite une séquence de huit conséquences
entre x8 and 0 pour les utilités suivantes: 0.015, 0.031, 0.062, 0.093, 0.125, 0.156, 0.187,
0.25, où U(Gr) = r en utilisant la probabilité pour les gains déterminée au début.

30
Le logiciel élicite également des données pour deux tests de cohérence. Pour le premier,
la différence d'utilité étant constante entre les éléments successifs dans les séquences
élicitées au début de l'expérience ({G0 , G1 , G2} pour les gains et {L0 , L1 , L2 } pour les
pertes), la cohérence exige que les égalités entre les différences d’utilité soient
maintenues pour les pertes et gains obtenus en étapes 2 et 4 de la procédure d'élicitation.
Le logiciel permet également un deuxième test de cohérence qui se compose de quatre
questions. Pour chacune des loteries (Gr, pr; Lr), r = 0.031, 0.0625, 0.125, 0.250, la
probabilité pr établie rend le gérant de portefeuilles indifférent entre la loterie d’une part
et ne rien recevoir d’une autre part. La même probabilité d'indifférence devrait être
obtenue dans les quatre questions.

Les Indifférences
La méthode de bisection est employée pour l'élicitation des indifférences à travers des
séries de choix. Le gérant de portefeuilles doit choisir entre deux loteries A = (1000, p;
600) et B = (L0, p; 100) qui correspondent à (L1, p’; L*) et (L0, p’; L) présentées comme
des “camemberts” sur l’écran de l’ordinateur où les morceaux correspondent aux
probabilités. Il doit décider pour chaque itération qui change la valeur de L0, s'il change
de la loterie B à la loterie A. A partir de L0 = 4000 (L0 peut changer de 4000 à 8000,
l'intervalle pour L0 étant fixé tel qu'il garantie une préférence forte pour B), le processus
consiste à rétrécir l'intervalle contenant L0 par un certain nombre d'itérations jusqu'à
obtenir la valeur de L0 qui rend le gérant de portefeuilles indifférent entre les deux jeux,

c’est à dire (1000, p; 600) ∼ (L0, p; 100). La conséquence L0 est ainsi déterminée par une
série d'itérations pour rendre le joueur indifférent entre les deux loteries. La méthode de
bisection employée pour produire les itérations est illustrée dans la Table 2 pour L1 and
L0.0625.

31
 Table 2: Illustration de la Méthode de Bisection

Iterations Choix Visant l’ Elicitation de L1 Choix Visant l’ Elicitation de L0.0625

1 (-1000, 0.33;-600) vs. (-4000,0.33;-100) -6080 vs. (-7800,0.78; 0)
2 (-1000, 0.33;-600) vs. (-2500,0.33;-100) -3040 vs. (-100000,0.78; 0)
3 (-1000, 0.33;-600) vs. (-1750,0.33;-100) -4560 vs. (-100000,0.78; 0)
4 (-1000, 0.33;-600) vs. (-2100,0.33;-100) -3800 vs. (-100000,0.78; 0)
5 (-1000, 0.33;-600) vs. (-2300,0.33;-100) -4180 vs. (-100000,0.78; 0)
Indifférence -2200 - 4370

La loterie choisie est en caractères typographiques gras et seule la conséquence qui est à
éliciter varie, c’est à dire augmente ou diminue en fonction du choix de l’itération. La
grandeur du changement est toujours la moitié du changement dans la question
précédente. Néanmoins, la conséquence qui en résulte doit être un multiple de 10 et la
probabilité résultante un multiple de 0,01, sinon, la valeur est fixée au plus proche
multiple de 10 ou de 0,01. La méthode résulte dans un intervalle dans lequel la valeur de
l’indifférence doit être choisie. Par exemple, la valeur de l’indifférence pour L0.062 doit
être entre −4180 et −4560 . Par conséquent la valeur est −4370 le mi-point de
l’intervalle. Au début des itérations, les valeurs sont généralement choisies de manière
que les lotteries aient des valeurs espérées égales. Les exceptions sont L1, L2, G1 et G2
dont les valeurs au début sont L0 +3000, L1 +3000, G0 +3000 and G1+3000.

ANALYSE DES DONNEES

Les données ont été analysées en trois étapes (présentées dans la section 4.3). D’abord, la
forme de la fonction de l'utilité pour les gains et pour les pertes pour chaque gérant de
portefeuilles a été déterminée en suivant l'évolution de la pente de la fonction d’utilité en
divers points (p: 166). Trois familles paramétriques ont été ensuite considérées pour
l’ajustement paramétrique: La famille puissance, la famille exponentielle et la famille
expo-puissance (p:167-168). Pour l’aversion aux pertes et dans l’absence d’une définition
généralement admise, le degré de l’aversion aux pertes des gérants de portefeuilles a été

32
mesuré suivant les définitions suivantes (p:169): Kahneman et Tversky (1979); Wakker
et Tversky (1993); Neilson (2002); Bowman et al. (1999); et Köbberling et Wakker
(2005).

Elicitation Non-Paramétrique de la Fonction d’Utilité

Pour pouvoir suivre l’évolution de la pente de la fonction d’utilité pour les gains et les
pertes en divers points, deux pertes élicitées Lr and Lr’ (Gr et Gr’ pour gains) sont définies
adjacentes si Lr > Lr’ (Gr > Gr’) et qu’il n’y a aucune perte (gain) élicitée dans l'intervalle.

S L− ( r ) est défini comme la pente du segment liant (Lr, U(Lr)) et (Lr’, U(Lr’)) où Lr and Lr’

sont adjacents. De même, S L+ (r ) est défini comme pente du segment liant (Lr”, U(Lr”)) et

(Lr, U(Lr)) où Lr” et Lr sont adjacents. SG− ( r ) et SG+ ( r ) sont définis de façon analogue.

∆S L ( r ) = S L+ ( r ) − S L− ( r ) dénote la variation de la pente autour de Lr quand on s’éloigne de

zéro. De même, ∆SG ( r ) = SG− ( r ) − SG+ ( r ) dénote la variation de la pente autour de Gr

quand on s’éloigne de zéro. On peut vérifier facilement que ∆Si (r ), i = G, L positif,

(négatif, zéro) correspond à l'utilité convexe (concave, linéaire). Sept valeurs de
∆SG (r ) ont été obtenues pour des gains et onze valeurs de ∆S L (r ) pour des pertes pour
chaque gérant de portefeuilles. Pour tenir compte de l'erreur de réponse, les gérants de
portefeuilles qui ont au moins quatre/sept ∆SG (r ) négatifs (positifs) ont été classifiés en
tant qu'ayant des utilités concaves (convexes) pour des gains. Ceux qui ont au moins
sept/onze ∆S L (r ) négatifs (positifs) ont été classifiés en tant qu'ayant des utilités
concaves (convexes) pour des pertes.

Ajustement Paramétrique
Les données ont été également analysées en supposant trois formes paramétriques: la
famille de puissance, la famille exponentielle et la famille expo-puissance. (Le lecteur
peut se reférer aussi à la section 4.3.3, p: 167-168 du document en anglais où les
équations relatives aux trois familles sont numérotées comme 4.6, 4.7 et 4.8
respectivement).

33
La famille puissance est définie avec α > 0 pour les gains et β > 0 pour les pertes
comme suit:
x
U = U max ( )α
if x ≥ 0 (1)
xmax
x
U = −( − ) β if x < 0
100000

Le cas où α < 0 correspond à une fonction d’utilité décroissante. La fonction puissance

pour les gains est concave si α < 1, linéaire si α = 1 et convexe if α > 1; pour les
pertes, elle est convexe si β < 1, linéaire si β = 1 et concave si β > 1.

La famille exponentielle est définie avec α ≠ 0 et for β ≠ 0 comme suit:

x
U max (exp(−α ( )) − 1)
xmax
U= if x ≥ 0 (2)
(exp(−α ) − 1)
x
(exp(− β ( )) − 1)
U =− 100000 if x < 0
(exp(− β ) − 1)

x x
La fonction est définie comme U max ( ) pour α = 0 , et comme − pour β = 0 ;
xmax 100000
la fonction pour les gains est concave si α > 0 et convex si α < 0; pour les pertes, elle est
convexe si β > 0 et concave si β < 0.

La famille expo-puissance a été introduite par Abdellaoui, Barrios et Wakker (2002).

C’est une variation d’une famille de deux paramètres proposée par Saha (1993). Elle est
définie comme suit avec pour α ≠ 0 et pour β ≠ 0 comme suit:
x
U max (1 − exp(−( )α / α ))
xmax
U= if x ≥ 0 (3)
1 − exp(−1/ α )
x
(1 − exp(−( ) β / β ))
U =− 100000 if x <0
1 − exp(−1/ β )

34
Les cas où α ≤ 0 et β ≤ 0 n’ont pas été considérés parce que contrairement à
l’observation ils mènent à une aversion extrême pour le risque pour les gains et une
tendance extrême au risque pour les pertes quand le zéro est parmi les conséquences de la
loterie. La famille expo-puissance pour les gains est concave si α ≤ 1 et convexe si α ≥
2; pour les pertes elle est convexe si β ≤ 1 et concave si β ≥ 2.

Mesure de L’Aversion aux Pertes

Plusieurs définitions de l'aversion pour les pertes ont été considérées: Kahneman et
Tversky (1979); Wakker et Tversky (1993); Neilson (2002); Bowman et al. (1999);
Köbberling et Wakker (2005). Excepté pour cette dernière localisée au point de référence,
elles sont toutes globales et certaines sont plus strictes que d’autres. Elles ont été
opérationalisées comme suit:

Kahneman and Tversky (1979) ayant suggéré que l’aversion pour les pertes devrait être
définie comme −U ( − x ) > U ( x ) for x > 0, le degré de l’aversion pour les pertes calculé
U (−Gr )
pour chaque gain élicité a été − pour tout G > 0 où r renvoie aux utilités
U (G r )
communes pour les pertes/gains: 0.015, 0.031, 0.06, 0.093, 0.125, 0.15, 0.18, et 0.25.

D'abord les U(–Gr) ont été calculés et ce en utilisant l'interpolation linéaire pour chaque
gain obtenu. Le nombre des conséquences obtenues dans le domaine des pertes étant plus
grand que le nombre de conséquences obtenues dans le domaine de gains, il a été possible
d'obtenir huit valeurs d' U(–Gr) pour chaque gérant de portefeuilles. Pour G0.25 excédant
100000, U(–Gr) a été pris en tant que -1.

Le gérant de portefeuilles a été classifié averse aux pertes quand au moins 6 des 8
valeurs des coefficients d'aversion aux pertes calculées avaient excédé 1, neutre pour les
pertes quand au moins 6 valeurs avaient été égales à un et chercheur de gain ou “gain
seeker” quand au moins 6 valeurs avaient été inférieures à 1. Le moyen/médian de
U (−G )
− a été pris ensuite comme le coefficient d'aversion aux pertes.
U (G )

35
 U (− x) U ( y )
Neilson (2002) ayant proposé de définir l'aversion aux pertes par ≥ pour
−x y
tout x et y positifs (il fournit aussi une fondation de préférence) un candidat possible est le
min(U ( Lr ) / Lr )
. Le portfolio manager a été classifié comme averse aux pertes quand le
max(U (Gr ) / Gr )

min(U ( Lr ) / Lr ) min(U (Gr ) / Gr )

avait excédé un et un chercheur de gain quand le
max(U (Gr ) / Gr ) max(U ( Lr ) / Lr )
avait excédé un.

Wakker et Tversky (1993) ayant défini l'aversion aux pertes comme U '( − x) ≥ U '( x) pour
U '(−Gr )
tout x > 0, le coefficient d’aversion considéré pour les pertes est où U '(Gr ) , la
U '(Gr )
pente ∆ Ur/ ∆ Gr a été calculée comme suit:

(U (Gr ) − U (Gr ' )) (U (Gr " ) − U (Gr ))

U '(Gr ) = 1/ 2[ + ] (4a)
(Gr − Gr ' ) (Gr " − Gr )

et

(U ( Ls ) − U ( Ls ' ))
U '( −Gr ) = (4b)
( Ls − Ls ' )

si Ls’ < –Gr < Ls et Ls et Ls’ sont adjacents, et en définissant,

(U ( Ls ) − U ( Ls ' )) (U ( Ls " ) − U ( Ls ))
U '( −Gr ) = 1/ 2[ + ] (4c)
( Ls − Ls ' ) ( Ls " − Ls )

si –Gr = Ls. (Le lecteur peut se reférer à la page 171 du document où les trois dernières
équations sont numérotées comme 4.9a, 4.9b et 4.9c respectivement).

Les pentes pour chaque gain ont été calculées d’abord, puis les pentes pour les pertes
commensurées suivant l’équation convenable. Sept valeurs ont été obtenues pour chaque
gérant de portefeuilles. Celui-ci a été classifié averse aux pertes quand au minimum

36
six/sept valeurs avaient excédé 1, neutre quand six/sept valeurs avaient été égales à 1 et
chercheur de gain quand au minimum six/sept valeurs avaient été inférieures à 1.

Une définition plus forte a été fournie cependant par Bowman et al. (1999): l'aversion aux
pertes tient si U ' ( − x ) ≥ U ' ( x ) pour tout x et y positifs. Autrement dit, la fonction d’utilité
pour les pertes est partout plus pentue que la fonction d’utilité pour les gains. Par
min(U '( Lr )
conséquent, le coefficient d'aversion aux pertes a été calculé comme
max U '(Gr )

min(U '(Gr )
excédant un et chercheur de gain comme excédant un. U '(Gr ) and U '( Lr )
max U '( Lr )
ont été calculés comme dans les équations pour les coefficients de Wakker et de Tversky
(1993).

Finalement, Köbberling et Wakker (2005) ont défini le coefficient d'aversion aux pertes
comme U '↑ (0) / U '↓ (0) où le numérateur et le dénominateur représentent respectivement

les dérivées gauches et droites de l'utilité au point de référence. Par conséquent, le

coefficient d'aversion aux pertes qui implique que la fonction d’utilité est plus pentue
pour de petites pertes que pour de petits gains a été calculé comme U(L0.015)/L0.015 sur
U(G0.015)/G0.015 c.-à-d. G0.015/L0.015. Le gérant de portefeuilles dont le coefficient avait
excédé 1 a été classifié comme averse aux pertes.

LES RESULTATS DE L’EXPERIENCE I.

Les résultats non-paramétriques (tabulés à la page 174, Table 11) montrent que le patron
le plus commun, la concavité pour les gains et la convexité pour les pertes exhibé par
58% des gérants de portefeuilles est plus grand que celui de Fenema et van Assen (1999),
Abdellaoui (2000) et Etchart-Vincent (2004) dont l’intervalle varie entre (37% et 47%) et
proche du 54% de Abdellaoui et al. (2006). Pour l’ajustement paramétrique ‹‹puissance››
(les résultats paramétriques sont tabulés à la page 176-177, Tables 12, 13, 13(a), le

37
coefficient médian pour les pertes (0,9) est cependant dans la marge des études
mentionnées ci-dessus qui varie entre (0,84) et (0,97). L'aversion aux pertes existe dans
l'agrégat avec une pente de 0,450 (R2 ajusté est 0,906). Quand aux coefficients d'aversion
aux pertes obtenus, ils varient avec les définitions utilisées soulignant le besoin d’une
définition généralement admise. A cet égard, et excepté pour la définition de Kahneman
et Tversky (1979), peu de gérants de portefeuilles ont été caractérisés selon les définitions
globales qui mesurent l’aversion aux pertes en divers points et qui semblent
excessivement fortes pour des buts empiriques. Les tables 14 et 15 (p: 178-181)
présentent les résultats de l’aversion aux pertes pour le gérant de portefeuilles
moyen/médian et les comparent aux résultats obtenus par Abdellaoui et al. (2006) pour
étudiants.

Contrastés avec les résultats des études précédentes pour étudiants (Schmidt et Traub,
2002; Abdellaoui et al., 2006) pour la définition de Wakker et Tversky (1993) le gérant
de portefeuilles médian est moins averse aux pertes (1,08) par rapport à l'étudiant
médian (1,43; 1,53) respectivement. Contrastés pour la définition de Kahneman et
Tversky (1979) avec Abdellaoui et al. (2006) et Bleichrodt et al. (2001) qui ont estimé un
coefficient d'aversion aux pertes selon la même définition les résultats médians sont par
contre proches de Abdellaoui et al. (2006) mais moindres que Bleichrodt et al. (2001).
(1,69 ; 1,72 ; 2,17 respectivement). Pour la définition locale de Köbberling et Wakker
(2005) le gérant de portefeuilles médian est plutôt non averse aux pertes avec un
coefficient de 0,74 par rapport à l’étudiant médian de Abdellaoui et al. (2006) dont le
coefficient est de 2,52 .

DISCUSSION DES RESULTATS

Les résultats non-paramétriques (discutés aux pages 184-188) indiquent donc que le
gérant de portefeuilles est moins averse que l’étudiant et que plus de gérants de
portefeuilles que d’étudiants sont agressifs dans des situations où ils doivent décider entre
une perte sûre et une perte probable.

38
Pour ce qui est de l’aversion aux pertes, les gérants de portefeuilles possèdent de par leur
activité professionnelle quotidienne une gamme de formation et un haut niveau de
connaissance qui font qu’il est plausible qu’ils évaluent les enjeux différemment des
étudiants. Certes, il faut aussi souligner que les entrevues ont été entreprises durant la
période (2003-2004) qui correspond à une croissance (l’index Standard et Poor est en
hausse de 28%, 17% et 10,9 % respectivement par rapport à l’année précédente) et il
n’est pas inconcevable que le degré d’aversion aux pertes des gérants de portefeuilles ait
diminué avec le mouvement en hauteur du marché boursier. Barberis, Huang and Santos
(2001) dont le modèle d’évaluation des biens est basé sur une aversion aux pertes qui
change dans un contexte dynamique n’auraient pas pu obtenir la volatilité observée dans
le marché boursier sans cette variation.4

Cette variation certes consiste en deux volets: des gains antérieurs rendent l’individu
moins averse aux pertes alors que des pertes qui succèdent à d’autres rendent l’individu
plus averse aux pertes (Thaler et Johnson, 1990; Gertner, 1993); aussi, est-il important de
clarifier que le deuxième volet ne contredit pas la convexité de la fonction d’utilité dans
le domaine des pertes mais seulement l’hypothèse d’intégration des biens. Comme le
remarquent Barberis, Huang et Santos (2001), si les investisseurs intégraients plusieurs
années de gains et de pertes, ils seraient en train d’évaluer des niveaux absolus de
richesse et non pas des changements dans la richesse. D’ailleurs, Thaler et Johnson
(1990) indiquent aussi que dans le cas d’une loterie qui résulte en conséquences limitant
la perte maximale mais offrant la chance de devenir quitte les étudiants sont enclins au
risque même après une perte.

Gross (1982) qui documente ce phénomène sur le marché boursier y refère comme la
“getevenitis disease”. Le dictat de la rationalité est douloureux: pour réaliser des pertes,
les gérants de portefeuilles doivent auparavant renoncer à l’espoir de s’en sortir quitte.

4
Le modèle de Barberis, Huang et Santos (2001) a été influencé par Kahneman et Tversky (1979) pour
l’aversion aux pertes and par Thaler et Johnson (1990) et Gertner (1993) pour la variation de cette dernière.

39
Finalement, il faut bien mentionner les compensations auxquelles le gérant de
portefeuilles aurait droit en cas de profits. A titre d’exemple, les primes de Wall Street
pour l’année 2006 ont varié de $1 million à $3 millions pour le directeur moyen, jusqu’à
$60 millions pour les maisons d’investissement comme Goldman Sacks, Lehman
Brothers et Morgan Stanley (New York Times, Dec 25, 2006).

Les différences de comportement des gérants de portefeuilles par rapport aux étudiants
pourraient donc être dues à des facteurs d’occupation. L’expérience II a été enterprise
pour analyser d’une manière qualitative les préférences des étudiants en MBA, gérants de
portefeuilles en potentiel en utilisant la méthode nouvellement développée de Baucells et
Heukamp (2006).

40
CHAPITRE V. EXPERIENCE II: EXPERIENCE AU LABORATOIRE
UTILISANT DES CONDITIONS DE DOMINANCE STOCHASTIQUE

La méthode de Baucells et Heukamp (2006) généralise les conditions de dominance

stochastique introduites initialement par Rothschild et Stiglitz (1970) et plus récemment
par Levy et Wiener (1998) et Levy et Levy (2002). En effet, les préférences peuvent être
inférées des choix entre des loteries construites de façon que l’une domine
stochastiquement l'autre. Un individu avec une fonction d’utilité non-décroissante et
concave c.-à-d. un individu averse au risque et qui maximize EU dis-préférera une loterie
qui est dominée par la dominance stochastique de deuxième ordre (SSD) et l'inverse est
également vrai: si l'individu élimine des alternatives dominées par SSD, sa fonction
d’utilité est concave. Pour distinguer entre les classes des fonctions d’utilité non-
décroissantes qui ne sont pas concaves partout, d'autres conditions sont certes
nécessaires. Levy et Levy (2002) développent la dominance stochastique des prospects
(PSD) et la dominance stochastique de Markowitz (MSD) pour différencier entre les
fonctions d’utilité qui sont concaves pour les gains et convexes pour des pertes (ayant une
forme de S comme celle postulée par Kahneman et Tversky (1992)) et ceux qui sont
convexes pour des gains et concaves pour des pertes (ont une forme de S inversé comme
celle postulée par Markowizt (1952)). Ces conditions peuvent être appliquées sous EU ou
sous n'importe quel modèle dépendant d’un point de référence qui n'incorpore certes pas
une fonction de transformation des probabilités.

Aussi, Baucells et Heukamp (2006) étendent-ils ces conditions à CPT en incorporant

cette dernière ainsi que l’aversion aux pertes et les utilisent comme guides pour concevoir
des paires de loteries ou tasks qui sont en compétition directe entre les deux théories
alternatives: CPT (1992) et Markowitz (1952). Ainsi, pour deux loteries F et G conçues
de facon que F domine G par la dominance stochastique de perspective ( F fPWSD G ) et G

domine F par la dominance stochastique de Markowitz ( G fMWSD F ) le choix de F (G)

implique que la fonction d’utilité du décideur a la forme S (S inversé). Les choix des
décideurs entre les loteries ainsi conçues indiquent des propriétés de leurs préférences et

41
leurs représentations sous un modèle de choix. Aussi, pour un modèle donné, la condition
de dominance stochastique liée à une caractéristique spécifique du modèle permet-elle
d'examiner les propriétés qualitatives de ce dernier.

L’expérience II utilise les paires de loteries construites par Baucells et Heukamp (2006)
pour examiner de facon qualitative les préférences des étudiants en MBA. Le chapitre V
se compose de six sections. La première présente l'intuition pour la dominance
stochastique de la théorie des prospects et de Markowitz suivies de la caractérisation des
préférences en utilisant ces conditions. La deuxième énonce l'objectif de l'expérience et
décrit la source de données. L'analyse des données est présentée dans la troisième section.
Les résultats concernant la forme de la fonction d’utilité sont présentés ensuite suivis des
résultats concernant l'aversion aux pertes et la probabilité globale des gains/pertes
respectivement. Une section finale discute les résultats de l’expérience.

DOMINANCE STOCHASTIQUE SOUS CPT ET MARKOWITZ

L'intuition en bref (élaborée à la page 192) est qu'un individu qui adhère à la théorie des
prospects donne plus d'importance aux conséquences près de l'origine qu'aux
conséquences extrêmes tandis que le contraire est vrai pour un individu qui adhère à la
théorie de Markowitz. Ainsi, au cas où aucune transformation des probabilités n’est
postulée, la différence en utilité espérée est donnée par:
b
∆ = ∫ [G (t ) − F (t )]U '(t )dt (5)
a

Où F et G sont les distributions cumulatives des loteries F et G et où l’on suppose pour

des raisons de simplicité que F et de G prennent les valeurs 0 et 1 pour certains a ≤ 0 et

b ≥ 0 respectivement. a et b correspondent aux deux points extrêmes d’inflexion dans la

fonction d’utilité de Markowitz et l’on s’attend à ce qu’ils indiquent les niveaux de
richesse extrêmes.

Comme l’indique l'équation (5) (c’est l’équation 5.1, p: 192 du document) la différence
[G(t) –F(t)] est mesurée en proportion avec U’(t); autrement dit, les segments où la pente

42
d' U(t) est large sont plus étirés relativement aux segments où la pente est faible. Les
secteurs près de l'origine sont magnifiés pour l’individu qui suit la théorie des prospects
puisque la pente est plus grande près de l'origine tandis que les extrémités sont
magnifiées pour l’individu qui suit la théorie de Markowizt puisque c'est là où la pente
est plus grande.

Cependant, quand la fonction de transformation des probabilités est factorisée dans le

processus de décision plus d'importance est donné aux conséquences extrêmes
relativement aux conséquences intermédiaires (les détails sont donnés à page 196). Ainsi,
quand cette fonction est factorisée dans l'équation (5), la différence d'utilité entre les
loteries F et G est donnée par
0 b
∆ = ∫ [ w(G (t )) − w( F (t ))]U '(t )dt + ∫ [ w(1 − F (t )) − w(1 − G (t ))]U '(t )dt (6)
a 0

En conséquence, en parallèle à l'étirage horizontal, l'axe vertical de probabilité cumulée

(0, 1) est étiré par la fonction de transformation des probabilités rendant les prospects
près de 0 (possibilité) et près de 1 (certitude) plus souhaitables et magnifiant la différence
entre F(t) et G(t) dans ces secteurs. Par conséquent, pour que CPT ne soit pas ambigue, il
est nécessaire de généraliser les conditions de dominance stochastique pour capturer cet
aspect important de CPT et également l’autre aspect de la théorie notamment l'aversion
aux pertes.

Pour généraliser ces conditions de manière qu’elles caractérisent les préférences d’un
individu qui adhère à CPT et les préférences d’un individu qui adhère à la théorie de
Markowitz, la fonction de transformation des probabilités sous CPT en forme de S
inversé est restreinte à un intervalle Wcd = Wc ∩ W d où Wc est la classe des fonctions de
transformation des probabilités convexes entre c et 1 et Wd la classe des fonctions de
transformation des probabilités concave entre 0 et d. c et d dénotent en même temps c+
et c- (ou d+ et d-) qui renvoient à w+ et w- les pondérations sous CPT pour les gains et les
pertes respectivement.

Aussi, UP est-elle définie comme la classe des fonctions d’utilité en forme de S telle que:

43
U ∈ UP si U ' ≥ 0 pour tout x ≠ 0, U " ≥ 0 pour x < 0 et U " ≤ 0 pour x > 0, et

UM est définie comme la classe des fonctions d’utilité en forme de S inversé telle que: U ∈
UM si U ' ≥ 0 pour tout x ≠ 0, U " ≥ 0 pour x > 0 et U " ≤ 0 pour x < 0.

Les conditions sont énoncées formellement comme suit:

Proposition 5.3:
− +
F fPWSD G si et seulement si F ≽ G pour tout U∈ UP, w − ∈ Wcd− , et w+ ∈ Wcd+ . De
facon similaire,
− +
F fMWSD G si et seulement si F ≽ G pour tout U ∈ UM , w− ∈ Wcd− et w+ ∈ Wcd+ ;

− +
F fSWSD G si et seulement si F ≽ G pour tout U ∈ Uconcave , w − ∈ Wcd− , et w+ ∈ Wcd+ ; et
− +
F fS *WSD G si et seulement si F ≽ G pour tout U ∈ Uconvexe , w− ∈ Wcd− , et w + ∈ Wcd+

Ainsi la fonction de transformation des probabilités incorporée est restreinte à c et d dont

le choix est le résultat d’un tradeoff. D’un côté, pour un c plus petit ou un d plus grand,
la portée des applications de ces conditions augmente mais d’un autre côté, un Wcd trop
restreint ne contiendrait peut-être pas les fonctions désirées. Les choix de Baucells et
Heukamp (2006) pour c et d ont été pris dans l’intevalle [0.05 to 0.88] déliéné par une
analyse de sensisivité. (Cette partie est traitée aux pages 197-200 où Figure 13 montre la
fonction de probabilité Wcd alors que les figures 12 et 14 montrent la distribution de
probabilités cumulative pour les tasks I et VII respectivement).

Pour la condition de dominance stochastique qui tient compte de la fonction de

transformation des probabilités et de l’aversion aux pertes, UL est définie comme la classe
des fonctions d’utilité telle que:
U ∈ UL si U '( − x) ≥ U '( x ) pour tout x ≥ 0, et UPL = UP ∩ UL, et la condition est

énoncée formellement comme suit:

44
Proposition 5.5:
− +
F fPWLSD G si et seulement si F ≽ G pour tout U ∈ UPL , w− ∈ Wcd− et w + ∈ Wcd+ tel que
s− ≥ s+

L’extension des conditions de dominance stochastique à CPT permet de tester: 1) la

courbe de la fonction d’utilité et/ou la présence de l'aversion aux pertes postulant une
fonction de transformation de probabilité en forme de S inversé ; 2) la courbe de cette
dernière postulant les caractéristiques empiriques pour la fonction d’utilité de CPT; 3)
finalement, si l’on postule que toutes les caractéristiques empiriques de CPT tiennent, une
violation de la condition de domaince stochastique implique une violation de CPT.

L'APPLICATION EXPÉRIMENTALE

Les loteries de Baucells et de Heukamp (2006) ont été utilisées pour examiner de façon
qualitative les préférences des étudiants en MBA à Arizona State University (ASU). Les
étudiants étaient au nombre de 40 et avaient été payés 10$ chacun pour compléter le
questionnaire. Plus spécifiquement, ils avaient à choisir pour 20 paires de loteries entre
deux investissements F et G introduits comme suit: “Supposez que vous avez décidé
d’investir 10000$ en action F ou en action G. Laquelle choisirez vous, F ou G quand il
est donné que le dollar gain ou perte dans un mois sera comme suit.” (Les paires de
loteries ou tasks se trouvent dans l’annexe D).

Comme les loteries F et G sont conçues de facon qu’elles aient la même valeur
mathématique espérée et que F fPWSD G et G fMWSD F le choix de F (G) implique le
rejet de la fonction d’utilité du décideur en forme de S inversé en faveur de celle en forme
de S. Le format du questionnaire le permettant, deux tasks ont été ajoutés pour étudier
l'impact de la probabilité globale de gains/pertes dans les jeux mixtes c’est à dire
comportant des gains et des pertes et les étudiants ont été également invités à commenter
sur les choix qu'ils ont faits.

45
ANALYSE DES DONNEES

Les 20 tasks utilisés dans cette expérience ont été initialement construits en réponse à la
déclaration de Levy et Levy (2002) que la théorie des prospects importe peu (is much ado
about nothing). En effet, les sujets de Levy et Levy (2002) qui avaient à choisir F ou G
pour trois tasks (I-III) où F fPSD G et G fMSD F avaient opté pour G dans les proportions
suivantes: 71%, 62% et 76% respectivement.

Aussi, les tasks I-III de Baucells and Heukamp (2006) imitent-ils les tasks de Levy and
Levy (2002) et discriminent entre l’hypothèse d’une fonction d’utilité en forme de S et
celle en forme de S inversé, le but étant d’obtenir précisément des réponses pareilles c’est
à dire U ∉ UP. Parce que F fPSD G , si w est linéaire (convexe) partout alors les choix

favorisant G impliquent que U ∉ UP .

Contrastés avec ces tasks (I-III), les tasks qui suivent (IV-VIII) discriminent entre
l’hypothèse d’une fonction d’utilité en forme de S et celle en forme de S inversé
factorisant cependant dans la décision la fonction de transformation des probabilités.
Ainsi, les loteries F et G sont des modifications des tasks I-III conçues de facon que
F fPWSD G et G fMWSD F avec w ∈ W0,10,4 , pour d ≤ 0,74 et c ≥ 0,1. Le choix de F c’est

à dire U ∈ UP combiné aves les résultats des tasks I-III implique que la fonction de
transformation des probabilités ne peut être linéaire ou convexe partout.

Pour spécifiquement examiner la courbe de cette dernière près de l’origine et postulant

que les caractéristiques empiriques de CPT pour la fonction d’utilité tiennent, le task IX
est construit comme une modification du task I avec un léger changement dans la
probabilité attachée aux conséquences communes. Dans F, un montant maximal a été
ajouté avec une probabilité de 2% et dans G un montant minimal a été ajouté avec une
probabilité de 2%. Conçu comme F fPWSD G and G fMWSD F ce task teste conjointement

avec le task I si l’hypothèse w ∈ W0.02 peut être rejetée.

46
Les quatre tasks qui suivent sont construits pour refuter l’argument que la concavité pour
les gains et la convexité pour les pertes sont poussées par l’effet de certitude. Aussi, les
tasks X-XI sont-ils construits comme des loteries toutes en gains et examinent la
convavité pour les gains et les tasks XII-XIII comme des loteries toutes en pertes et
examinent la convexité pour les pertes. Dans tous ces tasks aucune conséquence n’est
certaine et tous satisfont F fPWSD G pour c ≥ 0,1.

Quant à l’aversion aux pertes, sept tasks sont conçus pour l’examiner. Les tasks XIV-
XVI sont conçus comme F fW d LSD G et examinent uniquement l’aversion aux pertes sans
c

aucun postulat sur la fonction de l’utilité. Pour ces tasks, le choix de F implique donc
l’aversion aux pertes. Les tasks XVII-XX par contre sont conçus comme F fPW d LSD G et
c

examinent conjointement l’aversion pour les pertes et la concavité/convexité de la

fonction d’utilité. Pour les tasks XVII et XX le choix de F implique que les étudiants sont
averses aux pertes et que leur fonction d’utilité est concave pour les gains alors que le
choix de F dans les tasks XVIII et XIX implique que les étudiants sont averses auxs
pertes et que leur fonction d’utilité est convexe pour les pertes.

Finalement, les deux derniers tasks examinent l’effet de la probabilité globale des
gains/pertes. Les étudiants ont été présentés avec les loteries suivantes (Payne, 2005):
($100, 0.2; $50, 0.2; $0, 0.2; $-25, 0.2; $-50, 0.2). Ils devaient dans une première étape
ajouter (38$) à 0$ ou à 100$ et dans une deuxième étape ajouter (38$) à 50$ ou à 100$, le
choix d’ajouter (38$) à (0$) impliquant une préférence pour une plus grande probabilité
globale des gains. (Plus de détails sont donnés aux pages 204-205).

47
LES RESULTATS

Les tasks ont été donc divisés de facon que plusieurs hypothèses puissent être testées et
les résultats obtenus indiquent que:
1) pour la courbe de la fonction d’utilité, la majorité des étudiants a choisi F, choix qui
implique le rejet de la fonction d’utilité de l’étudiant en forme de S inversé en faveur de
celle en forme de S. En effet, pour les tasks IV-VIII comme prévu les étudiants ont choisi
F (67.5%, 85%, 87.5%, 85%, et 92.5%) respectivement. Ces proportions sont pareilles à
celles de Baucells et Heukamp (2006): 61%, 84%, 66%, 76% et 84% respectivement.
Pour les tasks X-XIII, les étudiants ont choisi F dans le proportions 85%, 87,5% 72,5 et
85% respectivement montrant que la concavité pour les gains et la convexité pour les
pertes ne sont pas poussées par l’effet de certitude. Pour les tasks IV- XIII les valeurs de
probabilité sont significatives et rejettent l’hypothèse nulle que % F = 0.5. (Ces valeurs se
trouvent dans la Table 20).

2) pour la fonction de transformation des probabilités, également comme prévu et

pareillement à Levy et Levy (2002) et à Baucells et Heukamp (2006), les étudiants ont
opté pour G pour les tasks I-III (62,5 %, 60%, 62,5% respectivement). Ainsi, le choix de
F c’est à dire U ∈ UP pour les tasks IV-VIII combiné avec le choix de G c’est à dire U ∉
UP dans les tasks I-III implique que la fonction de transformation des probabilités ne peut
être linéaire ou convexe partout.

3) pour le comportement de la fonction de transformation des probabilités près de

l’origine, le revirement des préférences de G à F pour les tasks I et IX (62,5% à 92,5%
respectivement) montre que w ∈ W0.02 est une classe plausible de ces fonctions et souligne
le changement brusque de la fonction près de l’origine, suggérant que les décideurs
utilisent l’intervalle des conséquences comme critère de décision.

4) pour ce qui est de l’aversion aux pertes, les résultats des tasks XIV-XVI conçus pour
déterminer uniquement l'aversion aux pertes montrent que les étudiants sont averses aux

48
pertes quand les probabilités sont les mêmes ou semblables. En effet, pour le task XIV,
60% des étudiants ont choisi F pour éviter une perte. La majorité de ceux qui ont opté
pour G n'a pas été attirée par la probabilité de 80% de ne rien gagner de F.

Pour le task XIV la fréquence de l'aversion aux pertes chez les étudiants à ASU (60%) est
plus grande que la fréquence (43%) chez les sujets de Baucells et de Heukamp (2006).
Ces derniers sont néanmoins composés d'étudiants et de professionnels qui ont choisi
différemment: la proportion des étudiants qui a choisi F comparée à celle des
professionnels était (48% vs 35%). Ceci supporte l’hypothèse que les professionnels sont
moins averses aux pertes que les étudiants. A cet égard, il convient de rappeler qu’une
comparaison des résultats de l'expérience I avec ceux d'Abdellaoui et al. (2006) a
également montré que l'étudiant médian est plus averse aux pertes que le professionnel
médian.

Le task XV où la perte maximale augmente (par rapport au task XIV), montre que
l'aversion aux pertes dans des loteries mixtes augmente à mesure que les enjeux
augmentent. En effet, la proportion d'étudiants qui ont choisi F (67.5%) a augmenté.
Certes, la probabilité de ne rien gagner de F a diminué de 80% à 60% (pour les tasks XIV
et XV respectivement).

Pour le task XVI qui examine l'aversion pour les pertes plus près de l'origine relativement
au task XV, moins d'étudiants (55%) ont préféré F à G préférant la plus grande
probabilité pour le gain le plus élevé dans G pour des probabilités globales égales de
gains/pertes.

5) les tasks XVII-XX sont un essai global pour CPT. Excepté pour le task XVII, qui
montre une préférence pour la probabilité globale de gain la plus élevée, F est clairement
préféré à G conformément avec les prévisions de CPT: 40%, 85%, 70%, 87,5%
respectivement. Pour le task XVII, 60% des étudiants ont déclaré avoir choisi G pour la
pour la grande probabilité globale de gain (70% dans G vs 50% en F) ou la probabilité

49
plus élevée du gain maximal et ainsi qu’ils se sont exprimés “500$, ce n'est pas beaucoup
à perdre”.

6) Finalement, Les deux derniers tasks, où 82%, 67.5% respectivement des étudiants ont
choisi F confirme que la probabilité globale de gains/pertes est un facteur décision
important. (Les tables 17, 18, 19, 20 p: 206-209) résument les résultats de l’expérience II
respectivement).

50
CONCLUSION

Pour généraliser les conditions de dominance stochastique de manière qu’elles

caractérisent les préférences d’un individu qui adhère à CPT, et les préférences d’un
individu qui adhère à la théorie de Markowitz, Baucells et Heukamp (2006) restreignent
la fonction de transformation de probabilité en forme de S inversé à un intervalle
Wcd = Wc ∩ W d . Le choix cependant de c/d est le résultat d’un tradeoff pris dans

l’intevalle [0.05 to 0.88] déliéné par une analyse de sensisivité et de l’avis même de
auteurs un désavantage par rapport à d’autres méthodes (Abdellaoui, 2000).

Leurs conditions tiennent aussi compte de l’aversion aux pertes telle qu’elle est définie
par Wakker et Tversky (1993). Ainsi, l’extension des conditions de dominance
stochastique à toute la théorie CPT permet de les utiliser comme guides pour concevoir
des paires de loteries qui sont en compétition directe entre CPT (1992) et Markowitz
(1952). Ainsi, pour deux loteries F et G conçues de facon que F fPWSD G and G fMWSD F
le choix de F implique le rejet de la fonction d’utilité du l’étudiant en forme de S inversé
en faveur de celle en forme de S. Alors que pour deux loteries F et G conçues de facon
que F fW d LSD G , le choix de F implique que l’ hypothèse que la fonction d’utilité est plus
c

pentue pour les pertes que pour les gains n’est pas rejetée.

En bref, les résultats de l'expérience II rejettent l'hypothèse que la fonction d’utilité

d’étudiants en MBA est convexe pour les gains et concave pour les pertes et sont
compatibles avec une fonction d’utilité en forme de S. En outre, ils réaffirment
l'importance de l’incorporation de la fonction de transformation des probabilités et
montrent que l’aversion aux pertes existe pourvu que les probabilités soient les mêmes ou
similaires. Cette condition est importante comme le souligne le résultat du task XVII, le
seul qui viole la troisième hypothèse (violation de CPT): 60% des étudiants ont déclaré
avoir choisi G soit pour la probabilité globale élevée de gain ou la probabilité élevée du
gain maximal combinée avec une perte extême limitée.

51
CONCLUSION GÉNÉRALE

La prise en compte des préférences de l'individu dans la prise de décision face à

l’incertain remonte au 18ème siècle où Cramer (1728) et Bernoulli (1738) ont proposé
indépendamment le modèle de l’utilité espérée. Axiomatizé deux cent ans plus tard par
vNM (1944) pour la prise de décision dans le risque et Savage (1954) et Anscombe et
Aumann (1963) pour la prise de décision dans l'incertain, ce modèle normatif est
aujourd’hui indéfendable comme modèle descriptif du comportement. En effet, les
travaux expérimentaux qui ont suivi les travaux théoriques ont montré que les gens
violent systématiquement certains axiomes de EU (e.g. Allais, 1953; Ellsberg, 1961;
Kahneman et Tversky, 1979). L'évidence empirique a motivé les chercheurs à développer
des théories alternatives de choix pour tenir compte des violations observées. Parmi ces
théories, CPT explique la plupart des violations et les phénomènes que EU est utilisée
pour expliquer grâce à l’intégration dans la théorie d’éléments non-normatifs notamment:
l'effet de réflexion, l’aversion aux pertes et le traitement subjectif des probabilités qui
découlent de la perception d’un point de référence. Tout en gardant le principe de
cohérence, elle raffine en plus ce qu’on comprend par l'attitude au risque en décomposant
le risque en trois facteurs: un fonction d’utilité, une fonction de probabilité et une
aversion aux pertes.

La popularité de CPT et la possibilité d'expliquer les anomalies sous EU par des

préférences averses aux pertes ont inspiré des économistes à rechercher des méthodes qui
peuvent éliciter/examiner (quantitativement/qualitativement) des fonctions d’utilité sous
CPT et tester l'aversion aux pertes et mesurer son degré. Ainsi, Abdellaoui et al. (2006)
élicitent sous CPT l’utilité pour les gains et pour les pertes simultanément pour mesurer
dans une seconde étape l’aversion aux pertes. Baucells et Heukamp (2006) étendent les
conditions de dominance stochastique à CPT. Ce travail expérimental utilise la première
pour éliciter les fonctions d’utilités des gérants de portefeuilles et la deuxième pour
inférer les préférences d’étudiants en MBA.

52
Les résultats de l'expérience I suivant Abdellaoui et al. (2006) montrent 1) que le patron
le plus commun est la concavité pour les gains et la convexité pour les pertes. En d’autres
termes plus de gérants de portefeuilles (58%) que d’étudiants (37% à 54%) sont
aggressifs en cas de pertes sûres vs des pertes probables. Pour l’ajustement paramétrique
“puissance” le coefficient médian pour les pertes est 0,9 (dans la marge des études
mentionnées ci-dessus), alors que pour l’exponentielle, le coefficient médian est 0,49; 2)
le gérant de portefeuilles médian est moins averse pour les pertes pour les définitions
globales alors que pour la définition locale de Köbberling et Wakker (2005) il est plutôt
non averse aux pertes: 0,74 vs 2,52 respectivement.

Néanmoins, Les résultats de l'expérience II suivant Baucells and Heukamp (2006)

confirment que pour la majorité des étudiants la fonction d’utilité en S inversé est rejetée
en faveur d’une fonction d’utilité en forme de S et que l’hypothèse de l’aversion aux
pertes ne l’est pas quand les probabilités sont les mêmes ou similaires. Pour l’aversion
aux pertes, la différence entre les fonctions d’utilité de l’étudiant médian et du gérant de
portefeuilles médian obtenues selon la méthode Abdellaoui et al. (2006) peut donc être
due à des facteurs de profession, notamment: la formation et le niveau de connaissance
acquis sur le terrain, la variation du degré d’aversion aux pertes avec le mouvement
récent (en hausse) du marché boursier, ainsi qu’aux compensations offertes en fin
d’année. Pour ce qui est de la convexité de la fonction d’utilité pour les pertes, il semble
qu’il est assez pénible aux gérants de portefeuilles de réaliser leurs pertes et d’admettre
avoir eu tort (les arguments sont présentés aux pages 212-216 du document).

53
REMARQUES ET DIRECTIONS POUR RECHERCHE FUTURE

Baucells and Heukamp (2006) proposent une nouvelle méthode aux expérimentateurs
intéréssés par la falsification d’une hypothèse particulière concernant la forme de la
fonction d’utilité ou la fonction de transformation des probabilités, sans avoir à l’éliciter.

Pour les expérimentateurs qui s’intéressent à l’exploration entière de la fonction d’utilité,

et la quantification du degré d’aversion aux pertes et son occurrence générale, Abdellaoui
et al. (2006) offrent une méthode non-paramétrique qui pourrait valider certaines
fonctionelles qui sont raisonnables. La méthode est basée sur l’élicitation de mipoints
d’utilité qui est souvent utilisée dans l’axiomatisation des modèles de décision et exige
moins de mesures en comparaison avec d’autres méthodes (Vind, 2003).

Combinés, les résultats de l'expérience I et ceux de l'expérience II indiquent l'importance

de la dépendance d’un point de référence en tant qu'élément de modélisation économique
comme préconisé par Rabin (1996) entre autres.

La convexité de la fonction d’utilité est un effet de dépendance d’un point de référence.

Comme c’est une tâche assez exigeante (Levy et Wiener (1998) notent que changer de
point de référence force l’investisseur à confronter et accepter ses pertes, ce qui est
douloureux) une utilisation corrective de la théorie des perspectives en analyse de
décision est conseillée ainsi qu’ a été suggérée par Bleichrodt, Pinto et Wakker (2001)
pour aider les gérants de portefeuilles à prendre de meilleures décisions dans leur interêt
aussi bien que celui de leurs clients.

L’aversion aux pertes est un autre effet de la dépendance d’un point de référence. Shalev
(2000) décrit l'existence de ce phénomène comme le résultat le plus saisissant des
fonctions d’utilité dépendantes d’un point de référence et prolonge l'analyse des jeux
pour inclure cette dépendance ainsi que l'aversion aux pertes. La première étape

54
cependant pour examiner ses prévisions est de mesurer l'aversion aux pertes au niveau
individuel. Pour une définition généralement admise de cette dernière, il faudrait peut-
être la démêler de la fonction de l'utilité. Pour les définitions globales, elle ne peut en être
séparée. Cependant, Köbberling et Wakker (2005) qui trouvent la séparation essentielle
pour la recherche sur les points de référence variables définissent l’aversion aux pertes
localement au point de référence avec une restriction qu’ils comptent néanmoins relaxer
dans une recherche future. Leur coefficient d’aversion aux pertes qui permet de classifier
tous les gérants de portefeuilles est soutenu par Schmidt and Zank (2002) qui
caractérisent l’aversion au risque sous l’hypothèse CPT par une condition jointe de la
fonction de l'utilité, de la fonction de la transformation des probabilités et de l’aversion
aux pertes. Le coefficient de Köbberling et Wakker (2005) est une première approche
axiomatique à l'aversion aux pertes en tant que composante logiquement indépendante de
l'attitude envers le risque. Néanmoins, il est à noter que si on compare les portfolio-
managers across les trois coefficients de Kahneman et Tversky (1979), Wakker et
Tversky (1993) d’une part et les coefficients de Köbberling et Wakker (2005) d’une autre
part, on trouve une grande différence.

Conçue pour décrire une population générale dans laquelle les différentes composantes
de l'attitude au risque (fonction d’utilité, fonction de probabilité et aversion aux pertes)
sont assez mixtes, CPT décrit un modèle naturel de réflexion et adresse les soucis des
gérants de portefeuilles d'une manière que ne peut le faire la normative EU. Pourtant les
théories qui décrivent le comportement des individus exhibant la prédominance extrême
d'une composante tel que l'EU et/ou la DT sont utiles parce que n'importe quelle position
extrême est plus claire et donc plus facilement reconnue et comprise que les positions
intermédiaires qui en aucune manière ne contiennent ou ne réconcilient les positions
extrêmes (Huxley, 1945). Dans ce sens, une théorie décrivant les individus qui exhibent
une prédominance extrême de l'aversion pour les pertes est utile.

55
TABLE OF CONTENTS

GENERAL INTRODUCTION 63

PART I: UTILITY THEORY AND DECISION MAKING 74

UNDER RISK

CHAPTER I. EXPECTED UTILITY THEORY 79

1.1 HISTORY OF THE CONCEPT OF UTILITY 79

1.1.1 How It Began 80
1.1.2 How it Evolved 81
Bentham 81
The Marginalist Revolution 81
The Ordinalist Revolution 82
The vNM Concept of Utility 83

1.2 EXPECTED UTILITY UNDER RISK 85

1.2.1 vNM Axiomatization’s: Axioms and Simple Probability 85
Measures
The General Framework 85
The Axioms 86
The Theorem 88

1.3 EXPECTED UTILTIY UNDER UNCERTAINTY 89

1.3.1 Savage’s Axiomatization: Axioms and States of the World 89

56
 The Axioms 90
The Theorem 91
1.3.2 Anscombe and Aumann’s: Axioms and Horse-Lotteries 93

CHAPTER II. VIOLATIONS OF EXPECTED UTILITY AND THE 96

ALTERNATIVE MODELS

2.1 VIOLATIONS OF CERTAIN AXIOMS OF EXPECTED UTILITY 96

2.1.1 Violation of the Axiom of Independence under Risk 97

The Common Consequence Effect 97
The Common Ratio Effect 99
The Transformation of the Probabilities 99
2.1.2 The Implication of the Violation for Dynamic Context 102
2.1.3 The Violation of the Axiom of Independence under Uncertainty 105

2.2 THE VIOLATION OF THE STABILITY OF PREFERENCES 106

2.2.1 Transitivity 107
2.2.2 Procedure Invariance 107
2.2.3 Descriptive Invariance 108
2.2.4 The Reflection Effect 108

2.3 THE ALTERNATIVE MODELS 109

2.3.1 Overview of Recent Developments in Utility Theory 109
2.3.2 Theoretical Predictions 111
The Probability Triangle 111
Expected Utility 112
Original Prospect Theory 113
Rank Dependent Utility 116
Dual Theory 120

57
CHAPTER III. CUMULATIVE PROSPECT THEORY 124

3.1 THE DESCRIPTIVE MODEL 124

3.1.1 The Non-Normative Elements in the Model 125

The Reference Point 125
The Localisation of the Reference Point 126
Loss Aversion 126

3.1.2 The Utility Function 128

3.1.3 The Weighting Function 129

3.2 THE FORMAL MODEL 131

3.2.1 Decision under Risk 131
3.2.2 Decision under Uncertainty 133

3.3 THE AXIOMATIZATION OF CPT 134

3.3.1 The Idea of Trade-off 134
3.3.2 The Intuition 135
3.3.3 The Axiomatization 136

3.4 THE CHARACTERIZATION OF RISK ATTITUDE 139

3.4.1 Notions of Risk Aversion 140
3.4.2 The Characterization of Risk Aversion under the Different Theories 143
Characterization of Risk Aversion under EU 143
Characterization of Risk Aversion under RDU 144
Characterization of Risk Aversion under CPT 145

CONCLUDING REMARKS FOR PART I 147

58
s
PART II: EXPERIMENTAL INVESTIGATION OF 150
INDIVIDUAL UTILITY FUNCTIONS AND LOSS AVERSION

CHAPTER IV. EXPERIMENT I: FIELD INVESTIGATION USING A 153

PARAMETER-FREE METHOD FOR THE ELICITATIONS

4.1 THE BACKGROUND 155

4.2 THE EXPERIMENT 156

4.2.1 The Elicitation Procedure 156

4.2.2 The Experimental Application 159

The General Set up 159
Details 161

4.3 DATA ANALYSIS 165

4.3.1 Consistency Tests for Non-Parametric Elicitation 165
4.3.2 Non-Parametric Elicitation of the Utility Function for Gains/ Losses 166
4.3.3 Parametric Fitting of the Data 167
4.3.4 Loss Aversion Measurement 169

4.4 RESULTS RELATED TO THE SHAPE OF THE UTILITY 172

FUNCTION

59
4.4.1 Consistency Results 173
4.4.2 Probability Weighting Tests 173
4.4.3 The Utility Function for Gains/Losses 173

4.5 RESULTS RELATED TO LOSS AVERSION 177

4.5.1 In The Aggregate 177
4.5.2 Individual Loss Aversion 177

CONCLUDING REMARKS FOR EXP. I 184

CHAPTER V. EXPERIMENT II: LABORATORY INVESTIGATION 184

USING STOCHASTIC DOMINANCE CONDITIONS

5.1 PROSPECT AND MARKOWITZ STOCHASTIC DOMINANCE 191

5.1.1 Levy and Levy’s (2002) Stochastic Dominance Criteria 191
The Intuition 192
Levy and Levy’s (2002) Preference Characterization 193

5.1.2 Baucells and Heukamp’s (2006) Stochastic Dominance Criteria 196

The Intuition 196
Accounting for the Probability Weighting 197
Accounting for Loss Aversion 200

5.2 EXPERIMENTAL APPLICATION 202

5.2.1 Objective 203

5.2.2 Source of Data 203

60
5.3 DATA ANALYSIS 204

5.4 RESULTATS RELATED TO THE SHAPE OF THE UTILITY 206

FUNCTION

5.5 RESULTATS RELATED TO LOSS AVERSION 207

5.6 RESULTATS RELATED TO THE OVERALL PROBABILITY OF 209

GAIN/LOSS

CONLUDING REMARKS FOR EXP. II 210

GENERAL CONCLUSION 212

REMARKS AND DIRECTIONS FOR FUTURE RESEARCH 217

BIBLIOGRAPHY 221

ANNEXE A: DATA FIGURES 241

Figure 8: The Certainty Equivalent for a Concave Utility Function 241

242
ANNEXE B: DISPLAY

DISPLAY 2 Illustration of the Use of the Scroll Bar 242

ANNEXE C: DATA TABLES 243

Table 5: Classification of Alternative Theories 243

Table 8: Practitioners’ Institutions and Locations 243

61
Table 10: Exponential and Power Risk Aversion Coefficients 244
Table 16: Main Parameters for Practitioners 245
247
ANNEXE D: THE QUESTIONNAIRE

LIST OF FIGURES 254

LIST OF TABLES 254

62
GENERAL INTRODUCTION

This research provides an experimental investigation of the fundamental preferences of

financial practitioners. The importance of the individual’s preferences in decision making
goes back in time to the 18th century when Cramer (1728) and Bernoulli (1738) proposed
independently that individuals do not maximize expected monetary values but expected
utilities, that is their subjective evaluation of the monetary values.

The expected utility model (EU) was not axiomatized however until 1944 by von
Neumann and Morgenstern for decision making under risk5, and a few years later by
Savage (1954) and Anscombe and Aumnann (1963) for decision making under
uncertainty.

The experimental works that followed the theoretical works showed however that people
violated systematically the axioms of EU (e.g. Allais, 1953; Ellsberg, 1961; Kahneman
and Tversky, 1979). The empirical evidence motivated researchers to develop alternative
theories of choice under risk and uncertainty to accommodate the observed violations.

These models termed “non-expected utility” or “generalizations of expected utility” were

tested subsequently in a probability triangle to find which if any best approximates actual
behavior (Harless and Camerer, 1994; Hey and Orme, 1994; Carbone and Hey, 1995).
The probability triangle is the set of probability distributions or lotteries that one can
define on a set of three different outcomes. This simplex is generally represented by the
unit triangle in a system of two rectangular axes. It is a useful device to display the
theories’ predictions. If the outcomes are fixed, the triangle contains the set of all possible
lotteries with the left edge characterized by a zero probability for the best consequence,

5
Bernoulli’s expected utility, Cramer’s and vNM’s have the same mathematical form but the first two are
assessed for certain outcomes and the third for risky outcomes.

63
the lower edge by a zero probability for the worst consequence and the hypotenuse by a
zero probability for the middle consequence.
None of the generalizations of expected utility however was found to fit the empirical
data significantly better than expected utility in the whole triangle, i.e. in all possible
risky situations. According to Abdellaoui and Munier (1998), the reason is that decision
makers preferences depend on the riskiness of the prospects they are facing, i.e. the
region of the probability triangle that represents their situation. A result of their
experiment is that non-expected utility models describe individuals’ preferences better
than EU outside the middle of the triangle. Starmer’s (2000) survey evaluates the data
from the triangle experiments and finds rank dependent utility models which do not
violate monotonicity as “probably the best bet” among the alternatives to expected utility
outside the interior of the triangle. In these models the probabilities are replaced by
decision weights that reflect the impact of events on the desirability of the outcomes.
These decision weights which result from weighting cumulative probabilities and sum to
one are allocated to the outcomes depending on their respective ranks in the series.

Among the rank-dependent models, prospect theory (Tversky and Kahneman, 1992)
based initially on “some pervasive effects people exhibit in their choices”, (Kahneman
and Tversky, 1979), also explains regularities of behavior in the field that are considered
paradoxes under EU. Camerer (1988) shows that prospect theory explains the anomalies
and the basic phenomena EU is used to explain in 10 fields because it integrates
psychological insights into economics.

SUPPORT FOR THE EXPERIMENTAL FINDING

Sketched in brief, the main features of prospect theory are: 1) the utility function defined
over gains and losses relative to a neutral reference point exhibits diminishing marginal
sensitivity to deviations from the reference point giving rise to an S shaped utility
function that is concave for gains ( u " < 0) and convex for losses ( u " > 0); 2) the utility
function is steeper in the domain of losses than in the domain of gains if the value of a
gain is larger in magnitude than the value of a commensurate loss, a characteristic labeled

64
by the authors as loss aversion; and 3) decisions are based on subjective cumulative
distributions given by π = w( p ) where p refers to the objective cumulative distribution, w
to a subjective treatment of p such that w '(.) > 0, w '(0) = 0 and w '(1) = 1. The
psychological insights integrated into the theory are hence: the reflection effect, loss
aversion and the subjective treatment of probabilities.

The reflection effect is a core psychological element of prospect theory. It refers to the
finding that the preferences between losses are “mirror images” of the preferences
between gains (Kahneman and Tversky, 1979, p: 268) that is, if individuals prefer sure
gains to probable gains, they would prefer probable losses to sure losses. The reflection
effect explains the reluctance to realize losses in the stock market. Terence Odean (1988)
finds that as Shefrin and Statman (1985) have predicted, investors realize their gains too
soon and are reluctant to realize their losses except for December for tax-motivated
selling and that their behavior has led them to lower returns. Graphically, investors would
have a concave utility for gains and a convex utility for losses provided the elicitation
method is robust against probability distortion (Abdellaoui et al., 2006). Previous
elicitations, using the trade-off (Wakker and Deneffe, 1996) which filters out probability
weighting, generally well corroborate the concavity for gains at the aggregate level and
for most individuals (Abdellaoui, 2000; Abdellaoui, 2002; Bleichrodt and Pinto, 2000).
The evidence for the convexity for losses is however, less clear-cut at the individual level
(Abdellaoui, 2000; Fenema and van Assen, 1999) and moreover, Levy and Levy (2002)
find a concave utility for losses and a convex utility for gains following Markowitz
(1952).

Another core element of prospect theory’s psychological findings is loss aversion. It

refers to the asymmetry of gains and losses relative to a reference point with losses
looming larger than commensurate gains. Kahneman and Tversky’s (1979, p: 279)
definition of loss aversion is as follows: An individual is loss averse if he dislikes
symmetric 50-50 bets and moreover, the aversiveness to such bets increases with the
absolute size of the stakes. Kahneman and Tversky show that the above behavioral
definition of loss aversion is equivalent under prospect theory to a utility function which

65
is steeper for losses than for gains. The latter result underlies the various definitions of
loss aversion that followed: (Tversky and Kahneman, 1992; Wakker and Tversky, 1993;
Bowman et al., 1999; Köbberling and Wakker, 2001; and Neilson, 2002).

The theoretical support for loss aversion comes from Rabin (2000) who shows that
plausible degrees of risk aversion over modest stakes under expected utility imply
unrealistically high degrees of risk aversion over large stakes while loss aversion implies
small-scale risk aversion and plausible risk aversion over large stakes. Rabin’s suggestion
favoring the use of loss aversion to explain risk aversion has been reiterated in
“Anomalies: Risk Aversion” by Rabin and Thaler (2001).

The empirical support for loss aversion comes from Camerer et al. (1997) for downward
labor supply and from Putler (1992) and Hardie, Johnson, and Fader (1993) for
asymmetric price elasticities. Loss aversion explains the endowment effect (Thaler, 1980;
Loewenstein and Adler, 1995), and thus the disparity between the willingness to pay and
the willingness to accept (Kahneman, Knetsch and Thaler, 1991; Bateman, Munro,
Rhodes, Starmer and Sugden, 1997), the status quo bias (Samuelson and Zeckhauser,
1988), and the disposition effect (Weber and Camerer, 1988, Odean, 1998; Heath,
Huddart and Lang, 1999). Benartzi and Thaler (1997) and Barberis, Huang and Santos
(2001) found loss aversion necessary to explain the large premium investors demand to
invest in stocks which was shown to imply an absurd degree of risk aversion under
expected utility by Mehra and Prescott (1985).

The possibility of explaining the anomalies under EU by loss averse preferences inspired
economists to look for methods that can test for loss aversion and measure its degree.
Experimental studies in the laboratory of measurement of loss aversion coefficients in the
aggregate include Fishburn and Kochenberger (1979); Tversky and Kahneman (1992);
Bleichrodt, Pinto and Wakker (2001); and Pennigs and Smidts, (2003). Studies that have
tested and measured loss aversion coefficients at the individual level are: Bleichrodt and
Pinto (2002); Schmidt and Traub (2002); Abdellaoui et al. (2006).

66
The third experimental finding integrated into the theory is the non-linear probability
weighting. In contrast to expected utility where the utility (of each possible outcome) is
weighted by its probability, in prospect theory the utility is multiplied by a decision
weight π (p) which is a monotonic function of p but is not a probability (Tversky and
Kahneman, 1986). The decision weight was introduced to model the tendency to
overweight low probabilities and underweight moderate and large probabilities. This
tendency was first observed by Preston and Baratta (1948). More recently, Diamond
(1988) and Edwards (1996) found that subjects judging low probability/high
consequence risks were more affected by the consequence size than by the probability
whereas those judging high probability/low consequence risk tended to combine the two.
According to MacCrimmon and Larsson (1979) subjects tend to choose the lottery with
the more probable gain for high probabilities of gain and tend to choose the lottery with
the highest gain for low probabilities of gain.

Further empirical support was later given by parametric studies that is assuming a
specific probability weighting function: Tversky and Kahneman (1992); Camerer and Ho
(1994); Tversky and Fox (1995); Wu and Gonzalez (1996); and by non-parametric
studies: Abdellaoui (2000) and Bleichrodt and Pinto (2000).

RESEARCH’S MOTIVATION

“I hope to show that much success has already been had applying prospect
theory to field data and to inspire economists and psychologists to spend more
time in the wild.”

Colin Camerer (Kahneman and Tversky, 2000, p: 288)

Until Abdellaoui et al. (2006) method, however, there was no method to elicit the utilities
for gains and losses simultaneously and non-parametrically in a relatively short time.
Wakker and Deneffe’s (1996) method does not assume any parametric function but elicits
the utilities for gains and losses separately, while Tversky and Kahneman’s (1992) and

67
Jullien and Salanie’s (2000) studies assume specific forms for utility. The availability of
the Abdellaoui et al. (2006) method and the ease of its applicability offered the possibility
to elicit non-parametrically the utility for gains and losses simultaneously and
consequently to measure the loss aversion/gain seeking coefficient of practitioners in the
stock market allowing at the same time a test of prospect theory in the latter field. The
elicitations required extensive traveling the interviews being conducted in the
practitioners’ offices at their respective financial institutions. However, the potential
benefits of eliciting the utility of financial practitioners and determining empirically their
parameters matter to economists interested in modeling behavior, to researchers of the
anomalies in the financial market and to decision analysts especially if they are
corroborated by further field investigations.

Eliciting the shape of the utility function provides a descriptive perspective to model
builders of a rational rule for decision making. Allais (1953; 1979) argues that rules of
how people should behave under uncertainty must take into consideration how
individuals actually behave.

Researchers interested in anomalies of the financial market would have available for their
studies parameters found independent of any assumption. For instance, regarding the
stock market, Benartzi and Thaler (1997) have used the Tversky and Kahneman (1992)
estimates of investor utility function and loss aversion coefficients to compute the
expected prospect values of stock and bonds in order to explain the equity premium
puzzle. Tversky and Kahneman’s (1992) method assumes a power form for the utility
function despite the drawback of confounding the general test of the theory to a specific
form, the estimation of all the parameters being problematic at that time (ibid, p:311).

Another benefit is in decision analysis where biased utilities can result in distorted
economic predictions. Traditionally, decision analysis assumes the normative expected
utility for calculating optimal decisions and for the elicitation of utilities. For the
assumption to be valid prescriptively however, the preferences of the decision maker
must be compatible with EU. The reluctance to realize losses evident in the practitioners’

68
elicited utilities as will be seen below implies that they find it hard to formulate their
decisions to sell a stock independently of its purchase price. The behavior violates EU but
is compatible with prospect theory. A corrective use of prospect theory has been
suggested by Kahneman and Tversky (1979); von Winterfeld and Edwards (1986);
Fischnoff (1991); and Kahneman and Tversky (2000, p: 157). Bleichrodt, Pinto and
Wakker (2001) propose the use of corrected utilities in prescriptions of optimal decisions,
the correction based on parameters preferably found empirically.

RESEARCH’S OBJECTIVES

The empirical investigation of the shape of utility function for gains and losses is the
primary objective of this experimental work which is divided in two parts: Experiment I
and Experiment II. Each experiment uses a different and newly developed method for
inferring preferences. Experiment I uses the Abdellaoui et al. (2006) method to elicit the
preferences of financial practitioners and Experiment II uses the Baucells and Heukamp
(2006) method to elicit the preferences of MBA students, potential future practitioners.

In Experiment I, financial practitioners’ utilities are elicited non-parametrically and

simultaneously for gains and losses. Parametric fittings to the power, exponential and the
expo-power families are undertaken to find out with which family the non-parametric
findings agree. The latter could validate reasonable functional forms and thus the use of
certain types of parametric estimation procedures which have the advantage of smoothing
response errors.

Abdellaoui et al.’s (2006) method also allows the measurement of loss aversion in the
aggregate and at the individual level. There is not however a commonly accepted
definition of loss aversion. The method nevertheless is definition-free and allows the
measurement of loss aversion at the individual level under both global and local
definitions. Abdellaoui et al.’s (2006) method can test thus prospect theory’s basic tenets

69
regarding the utility function referred to as value function in prospect theory’s
framework.

The results in the field for the shape of the utility function and for loss aversion are then
contrasted with the results of Abdellaoui et al. (2006) in the laboratory to examine
differences in behavior if any between practitioners and students.

Baucells and Heukamp (2006) having also made available a new method6, Experiment II
investigates the preferences of MBA students, potential financial practitioners and tests
for loss aversion. More specifically, Experiment II investigates the shape of the utility
function according to stochastic dominance criteria Baucells and Heukamp (2006) have
newly designed. The latter were motivated in developing them by Levy and Levy (2002)
who had found an inverse S shaped utility function following Markowitz (1952).
Baucells and Heukamp’s (2006) experiments using lotteries designed on the above
criteria reject the S shape and find general evidence for loss aversion. In Experiment II,
MBA students at Arizona State University (ASU) were asked in groups of 4-5 to respond
to the 20 tasks constructed by Baucells and Heukamp (2006). The latter having
hypothesized the impact of the overall probability in mixed gambles, Experiment II has
also included tests of this effect and the format of the tasks allowing it, the students were
also asked to comment on their choices in order to understand their reasons for any shift
in behavior that might occur as in between gain seeking and loss aversion for instance.

6
The idea of undertaking Experiment II came when Baucells, M. presented the paper he and Heukamp, H.
have co-authored at the FUR XI-Paris 2004 in France which I attended. Experiment II was thus
implemented based on the paper presented at that time and not on the reviewed paper that was published in
2006.

70
GENERAL OUTLINE

Both experiments I and II aim at inferring individuals’ preferences from the choices
presented to them. The choices were constructed within the framework of individual
decision making theory under risk. This work is thus divided in two parts: Part I presents
the theoretical framework of decision making under risk and Part II the experimental
work.

Part I consists of an introduction and four chapters. The introduction situates the major
milestones in the history of decision theory up to the early 20th century representation of
preferences which was derived under certainty. The distinction between risk and
uncertainty is introduced and narrows the theoretical part to the representation of
preferences in situations of risk. The prospects presented to the practitioners and to the
MBA students to infer their preferences were designed for situations under risk a special
case of uncertainty. Consequently, the elicited preferences of the financial practitioners
and the MBA students are approximations of their true preferences only under risk. The
concentration on risk as opposed to uncertainty will be observed for all four chapters
pertaining to the theoretical framework.7

Chapter I presents thus expected utility the first individual decision making theory for
risk with a brief description of subjective expected utility, the model being a normative
theory for both risk and uncertainty. The two models share the same bilinear form and
have similar axioms the main difference between the two is the representation of beliefs.

Chapter II explores the violations of the normative axioms of expected utility that dictate
how people should behave and presents the alternative models of decision making

7
Camerer and Weber (1992), Karni and Schmeidler, (1991a, part 3) and Fishburn (1988, ch. 7- 9) provide
reviews of developments in modeling preferences under uncertainty.

71
designed to have more descriptive power yet retain “desirable” criteria of rationality like
transitivity and monotonicity (Quiggin, 1982).

Chapter III is devoted to the theory that has had “much success in the field”, cumulative
prospect theory (1992), a review of the original prospect theory (1979). The chapter
explores the model’s psychological elements before presenting the formal model and the
axiomatization of the model under risk. The last section is concerned with the
characterization of risk aversion under the different theories. The notions of risk aversion
defined independently of any model are presented first to establish in a second step the
links between these notions and the preference functions under the different theories.
This final section shows that the understanding of risk aversion has been refined to a
large extent in the alternative theories frameworks’ challenging as a consequence the role
of utility in representing at the same time attitude towards consequences and attitude
towards beliefs. Thus, observed risk behaviors of different nature which are not
distinguished under expected utility are separated under the alternative models into
independent components. Moreover, prospect theory’s empirical realism further refines
the understanding of risk aversion and argues for its use in applied economics at least in
the specific context of portfolio selection the choice domain of financial practitioners and
MBA students.

Part II presents the experimental work and consists of two experiments. To each, a
chapter is devoted which includes a brief introduction showing the gradual yet portentous
development of the elicitation method from its predecessors, the elicitation procedure per
se, the data analysis and the results.

In chapter IV the utilities of practitioners are elicited non-parametrically following the

Abdellaoui et al.’s (2006) method and their loss aversion coefficients measured. Section 1
presents the parameter-free method. Section 2 is concerned with experimental
application. Section 3 presents the data analysis which includes the non-parametric
elicitations, the parametric fitting and the measurement of loss aversion in the aggregate
as well as the individual level. Section 4 and 5 present the results related to the

72
practitioners’ shape of the utility function and their loss aversion respectively. A final
section discusses the results and concludes.

In chapter V the shape of utility functions of MBA students are inferred from lotteries
based on stochastic dominance criteria developed by Baucells and Heukamp (2006).
Section 1 presents the intuition for clarification for stochastic dominance conditions
followed by the characterization of preferences using these criteria; section 2, the
experiment per se that is the objective, and the source of data; section 3 consists of the
analysis of data and section 4 and 5 give the results pertaining to the shape of the utility
function and loss aversion of MBA students respectively; section 6 investigates the
impact of the overall probability of gain/loss in mixed gambles. A final section discusses
the results and concludes.

The general conclusion is followed by some remarks and directions for future research.
The bibliography and the Appendix are presented next.

73
PART I: UTILITY THEORY AND DECISION MAKING
UNDER RISK

Part I of this research sets the theoretical scene for the experiments in Part II in which the
fundamental preferences of financial practitioners and MBA students in situations of risk
are elicited. Elicitation procedures presuppose the individual to decide among given
alternatives as if he is optimizing some underlying preference function. Generally, the
standard preference function in the elicitation procedure results from expected utility the
normative model of choice which describes how rational agents ought to choose.

However, for a normative model to be operational and prescriptively useful in

applications designed to aid decision makers, the actual behavior of an individual in
simple choice settings must be compatible with the behavior assumed in the model. In
other words, expected utility must be also defensible as a descriptive model of the
behavior of unaided decision makers otherwise, assuming expected utility to elicit the
individual’s preferences and attitudes towards risk cannot be meaningful.

Expected utility has been found however to be violated systematically in experimental

works. The earlier violations found by Allais (1953), Ellsberg (1961) and Kahneman and
Tversky (1979), stimulated the research for an alternative model normatively attractive
but with more descriptive power to accommodate the violations. The research does not
always cohere however because the researchers’ purposes are different some focusing on
the theories per se, that is on the mathematical properties of their axioms, others on their
descriptive validity, others yet on their implications in the field (Camerer and Weber,
1992).

Nevertheless, reviews of the alternative models (Camerer, 1992; Schmidt, 2002; and
Starmer, 2000) which organize the data from a large amount of research show a number
of stylized facts across the various studies that not only promote what has been called

74
since ancient times a hedgehog8 perspective but also are key ingredients in the selection
of the theory that might show, when tested in Experiments I and II, a good approximation
of the underlying preference function the individual is assumed to optimize. For if the
measurement of the model’s postulated parameters or functions in the laboratory and/or
the field shows the approximation is good, the model is approximately true and useful
despite unrealistic axioms (Camerer and Weber, 1992).

In hindsight, the violations of the expected utility properties which are violations of its
axioms can be categorized under two broad headings: those that violate the form of the
preference function and therefore violate only the independence axiom responsible for
restricting it strongly and those that violate the existence of a real-valued continuous
preference function and therefore challenge the axioms of ordering and continuity.

The alternative models that were developed in response to the violations are categorized
accordingly: Those that can be expressed in terms of a single preference function but
generalize expected utility by weakening the independence axiom and those that cannot
be reduced to a single function. The former models fall under the category of the so-
called conventional approach, and the latter fall under the non-conventional approach
Starmer (2000). One of these alternative models is however a cross fertilization of the
two strategies. Prospect theory (1992) assumes a single preference function, explains the
violation of the independence axiom and the strong empirical evidence for pervasive
phenomena like loss aversion and the reflection effect, which are inconsistent with an
evaluation in terms of final wealth as in the conventional approach.

Part I reviews the normative base theory expected utility and the theories designed as
alternatives in an attempt to find the model that can as much as possible reconciles
rational assumptions with experimental facts in order to be usefully assumed in
experimental elicitations. Expected utility is the natural point of departure for

8
“The fox knows many things, but the hedgehog knows one big thing” (Ignatieff, 1998, the Greek classical
poet Archilochus)

75
understanding the alternatives since they are generalizations of this standard theory. The
review is however narrowed in two significant respects:

First, the focus is on modeling choice under risk as opposed to the more general modeling
under uncertainty since the objective of Part I is to present the theoretical framework that
corresponds appropriately to the empirical tests presented in Part II and the prospects
presented to the practitioners and to the MBA students to infer their preferences were
designed for situations of risk.

To distinguish between risk and uncertainty is to distinguish between whether the

probability, “uncertainty’s yardstick”, (Fishburn, 1970, p: 101) is known or unknown.
The decision maker is in a situation of risk if each action leads to one of a set of possible
specific outcomes, each outcome occurring with an objective probability, agreed-upon
and impersonal. Certainty is a case of risk where the probabilities are 0 or 1.

The decision making is in a situation of uncertainty if each action leads to one of possible
specific outcomes, the probability of which is at best subjective (known ambiguously) or
at worst indeterminate.9 Risk is a case of uncertainty where the probability is known
unambiguously.

Second, the review of the alternative theories to expected utility under risk concentrates
on models presupposing a single preference function, yet defensible as a descriptive
models of actual behavior. The rationale for keeping this assumption is that it is an
important tenet of coherence and it’s not unreasonable to assume that people wish to obey
it even if it is a demanding task. Another is that although, the empirical violations of a
single preference function do make a case for the non-conventional models when these
are not judged using rational criteria, abandoning the notion of well-defined preferences

9
Subjective probability refers to a personal degree of belief as opposed to an impersonal, agreed-upon
degree of belief. Under expected utility it’s always known (inferred from bets), however, Ellsberg (1961)
showed that because of missing information regarding “the amount, type, reliability and unanimity of
information” individuals do not treat subjective probabilities as objective probabilities. In that case,
subjective probability is known ambiguously. Camerer and Weber (1992) discuss ambiguity in length.

76
requires changes that increase the complexity of the theory, reduce its predictive yield
and render it less compatible with the rest of economic theory10 (Starmer, 2000). In
Arrow’s (1995) words, “these models are apt to be very correct, it’s just their predictions
are a lot more vague than those implied by rationality; rationality is unique.”

Part I, has three chapters and is organized as follows:

Chapter I presents at first the history of the concept of utility which situates the major
milestones up to its modern form, the von Neumann and Morgenstern expected utility
theory (1944). The axiomatization of this first theory for individual decision making
under risk is then presented in terms of simple probability measures following Fishburn
(1970). The more general subjective expected utility is presented next showing risk as a
case of uncertainty and finally a few words on the directions expected utility theory under
risk took in the years that followed its axiomatization.

Chapter II provides 1) an overview of the violations of the properties of EU which are

violations of its axioms and 2) sets out the alternative theories that can account for the
descriptive invalidity of expected utility while retaining the principle of coherence. For
the latter, chapter II builds on previous overviews (Camerer, 1992; Schmidt, 2002) and
particularly on Starmer’s (2000) which evaluates the alternatives against empirical
evidence and finds the rank dependent weighting models to be “probably the best bet”
(Starmer, 2000, p: 359).

Chapter III explores the different aspects of the most popular among the many alternative
models constructed, prospect theory (1992) which in addition to retaining desirable
criteria of rationality like transitivity and monotonicity provides a convenient way of
modeling the influence of pervasive phenomena like loss aversion and the reflection
effect on choice. The last section of the chapter is devoted to the modeling of risk
preferences under the different utility theories, the shape of the utility function elicited
implying different risk attitudes depending on the model assumed. The equivalence, for

10
Tversky and Kahneman (1991) and Kahneman, Knetch and Thaler (1991) argue however for abandoning
the notion of stable preferences in favor of preferences indexed to a reference level which can be located
for particular cases.

77
instance, of the convexity of the utility function to risk seeking under EU does not hold
under prospect theory because of the non-linearity of the probability function in the latter.
The section begins with the notions of risk aversion defined behaviorally and
characterizes subsequently risk aversion under the different theories.

78
CHAPTER I. EXPECTED UTILITY THEORY.

Chapter I is devoted to the normative theory of decision making. It consists of three

sections. Section 1.1 narrates the earlier phases in utility theory up to the middle of the
twentieth century. The axiomatizations of expected utility under risk and under
uncertainty are presented in section 1.2 and 1.3 respectively. The final section reviews
briefly the experimental works which followed the theoretical works.

1.1 THE HISTORY OF THE CONCEPT OF UTILITY

“To change shape is in the very nature of history, because it is in the nature of
history to go on adding to itself”.

Arnold Toynbee (1972, p: 13)

History has added 200 years to its length between Bernoulli’s (1738) proposal of
expected utility maximization and von Neumann and Morgenstern’s (1944) modern
concept of utility and each portentous addition changed the whole. To establish a
common language, this section reviews the major milestones in the history of utility
theory during that period. The first is Daniel Bernoulli’s (1738) proposal of a theory of
expected utility as a basis for a decision making under risk using a logarithmic function
of wealth. The embracement of the utility concept by Jeremy Bentham (1789) in an
attempt to establish it as the basis of social policy constitutes the second.

The principle of diminishing marginal utility and the relationship between demand and
utility were not established however until the marginalist revolution in the 1870’s. This
third high point in the history of utility theory is soon followed by the ordinalist counter-
revolution whereby the utility under certainty and the cardinal comparisons were

79
abandoned in favor of an ordinalist vision of utility in which the principle of diminishing
marginal utility is nevertheless implicit.

The von Neumann and Morgenstern’s (1944) landmark constitutes the modern concept of
expected utility. von Neumann and Morgenstern were not however, the first to
incorporate explicitly uncertainty in the preference structure. Earlier, Ramsey (1931) has
constructed the first operational model of subjective expected utility in which preferences
are represented formally in terms of a utility function and a probability function and
Savage’s (1954) is a complete axiomatization of this earlier model. The von Newmann
and Morgenstern (1944) and the Savage (1954) axiomatizations will be presented in the
subsequent sections.

1.1.1 How It Began

The concept of expected utility maximization was introduced by Bernoulli (1738) and
Cramer (1728) independently in response to the inadequacy of the expected value
model11 to evaluate a game devised by Bernoulli’s cousin the so-called the St-Petersburg
puzzle: A fair coin is tossed until heads appears. The player receives 2 n if the first head
appears on trial n whose probability of occurrence is (1/2)n. What price is the player
expected to pay to enter the game?

The puzzle is that the expected value of the gamble is infinite as can be seen:
1 1 1
Expected Value = * 2 + * 4 + *8 + ... = 1 + 1 + 1 + …= + ∞
2 4 8

Yet most people would find it reasonable to pay a relatively small amount to play.
Bernoulli suggested that people didn’t maximize expected values but expected utilities,
that is, they averaged their subjective evaluations of the monetary outcomes rather than
the monetary outcomes themselves: “the determination of the value of an item must not
be based on its price but rather on the utility it yields. The price of an item is dependent
11
Probability theory and its application to problems of gambling were already highly evolved in the 18th
century. The evaluation of a game by the expected value decision rule goes back in time to Pascal (1623-
1662) and Fermat (1601-1665).

80
only on the thing itself and is equal for everyone; the utility however is dependent on the
particular circumstances of the person making the estimate.” (Bernoulli, 1738, p: 24).

Bernoulli posited utility as a function of wealth that increases with wealth at a decreasing
rate: “the utility resulting from an increase in wealth will be inversely proportionate to the
quantity of goods previously possessed.” The St Petersbourg’s expected utility with U a
logarithmic function of the monetary outcome is equal to 2.9 a reasonable price.
Bernoulli’s choice however of a logarithmic function, was “ad hoc”.12

1.1.2 How It Evolved

The concept of utility didn’t evolve till 50 years later with Jeremy Bentham (1789) who
founded utilitarianism. The principle of diminishing marginal utility was not recognized
however till the marginalists’ revolution in the 19th century and is implicit in the counter-
revolution of the ordinalists that soon followed. The expected utility model intuited by
Bernoulli and Cramer 200 years earlier in response to the inadequacy of the expected
value model was not axiomatized till the pioneer work of vNM (1944).

Bentham
Utility is referred to fifty years later by Jeremy Bentham (1789) as the pleasure or relief
from pain associated with the consumption of a good or a commodity and becomes the
foundation of utilitarianism, a theory of social choice in which Bentham advocates
aggregating individual utilities into one total utility and maximizing it, to achieve “the
greatest good for the greatest number”.13

The Marginalist Revolution

However, the insight that the principle of diminishing marginal utility rather than total
utility is the basis for engaging in commodity exchange was not perceived until the 19th

12
Cramer’s choice function was the square root function.
13
Although compatible with Adam Smith’s (1776) invisible hand (In every human breast...self-regarding
interest is predominant over social interest” (Bentham, The Book of Fallacies, 1824)) the utility concept
was not retained in 18th century’s classical economics.

81
century with the marginalist revolution associated with Jevons (1871), Menger (1871)
and Walras (1874). The marginalists were the first to recognize that the value of a
commodity depends on the demand for it,14 and that the demand for a commodity
depends on the utility associated with consuming the last unit of the commodity and not
on the total utility associated with consuming the commodity. A relationship between
demand and utility is thus established (Quiggin, 2004). A rational consumer in an
economy where prices are given should make the necessary exchanges (purchases and
sales) to move to a new endowment at which the ratio of his marginal utilities for
commodities equals the corresponding ratio of their prices. The grand rational
optimization of total utility in Bentham’s theory becomes a problem of individual rational
optimization. For the first time, the macro vision of the economy shifts to a micro vision
and the preferences (tastes) of the individual, the consumer, start to play an important role
in economic analysis.

The Ordinalist Revolution

However, the marginalists’ concept of utility as well as Bernoulli’s early statement of
expected utility theory and Bentham’s utilitarianism assume the existence of a cardinal
utility15 that is numerically measurable. The difficulty of constructing a numerical
measure of the unobservable concept of utility and of comparing utility scales between
individuals on one hand, and the development of the indifference curves analysis, by
Edgeworth (1881) and Pareto (1906) as a tool to measure preferences, on the other hand,
led to the rejection of the concept of a cardinal utility as a basis for constructing
individual preferences.

The notion of preference as a psychologically primitive concept was emphasized instead.

The consumer is assumed to be always able to state which of two alternatives he prefers
or else state that he is precisely indifferent between them. If his preferences are also

14
As opposed to the classical view of value deriving from production and distribution.
15
A cardinal utility is an ordinal utility (x ≽ y ⇔ u (x) ≥ u (y)), which is unique up to a positive linear
transformation.

82
transitive it follows from this set of simple axioms, that they can be represented by a real-
valued utility function that is preferred choices have higher utility numbers. Since only
the ordering property of the utility numbers is meaningful the utility function representing
the individual’s preferences is said to be ordinal. Preferences represented by indifference
curves16 requiring only the use of ordinal utilities that rank commodity bundles but does
not compare the differences between bundles were sufficient for the purposes of demand
theory.

A rational consumer then should make the necessary exchanges to move to a new
endowment at which the marginal rates of substitution between commodities, which
leave total utility unchanged and are observable, replace the ratio of his marginal utilities
for commodities.

However, as Quiggin (2004) argues, demand functions can only be well-defined if

preferences over commodity bundles are convex that is if a bundle containing an
appropriate mixture of two goods is preferred to either of two equally valued bundles
each containing only one of the goods. Implicit in this kind of convexity is the principle
of diminishing marginal utility.17 Convexity captures the intuition nevertheless by
referring to observable preferences rather than to unobservable utilities.

The vNM Concept of Utility

By the mid 20th century however, the axiomatic method “de rigueur” in mathematics
around the turn of the century, was embraced by other disciplines that relied on
mathematical methods (Nau, 2004). von Neumann and Morgenstern (1944) moving away
from their predecessors’ employment of indifference curves for rational optimization
used the axiomatic method to deduce formally that an individual strives to maximize his
utility. The expected utility model intuited by Bernoulli and Cramer 200 years earlier in
16
Indifference curves are taken to indicate corresponding trade-offs of goods A for B or B for A over the
same interval, the movement in either direction presumed to be the same. The reversibility of indifference
curves a central assumption of demand theory under certainty implies indifference curves do not cross.
If x≽ z and y≽ z then α x + (1−α ) y f z for α ∈]0,1[ . In particular, if x and y are indifferent to each other,
17

then α x + (1−α ) y f z is strictly preferred to either of them. The statement captures the intuition of
diminishing marginal utility namely that ½ x provides “more than half as much utility as x " (Nau, 2004).

83
response to the inadequacy of the expected value model is in the work of vNM (1944)
deduced from a relatively small set of axioms about rational preferences.18

For the rigorous description of the endeavor of the rational individual to maximize his
utility, vNM (1944) resurrected cardinal utility and axiomatized its maximization over
probability distributions under risk. Quiggin’s (2004) pertinent remark “no sooner driven
out of the front door of economic theory, cardinal utility re-entered it through the back
gate of game theory and expected utility theory” illustrates the immediate rekindling of
the debate of ordinal vs cardinal utility, the debate ending finally in an agreement: the
vNM’s utility associated with risk is different from the utility under certainty of the
marginalists (Baumol, 1958; Fishburn, 1989).19 Under vNM’s expected utility theory, the
individual is assumed to be able to compare different consequences and different
1 1
combinations of consequences. Thus, if he prefers z ≻ x ≻ y and x f ( y , ; z , ) then, it
2 2
can be inferred that his preference of x over y exceeds his preference of z over x and
differences in utilities become numerically measurable (vNM, 1944).

The vNM expected utility theory was not however the first to incorporate explicitly
uncertainty in the preference structure. Ramsey, a British philosopher and mathematician
to whom the theory (that a person’s actions are completely determined by his desires and
opinions) seemed a useful approximation of the truth, (Ramsey, 1931, p: 75) constructed
the first operational model of expected utility in which desires and opinions are
quantified and preferences represented formally in terms of a utility function and a

18
The expected utility model today stands on its own, but originally it was designed as part of vNM’s
modeling of rational social behavior in the playing of games, the simplest setting in which human
rationality is exercised, portraying for the first time decisions as reactions to others actions’ rather than
reactions to exogenous prices (Herbert Simon’s review of vNM’s book, p: 559). Rationality is assumed in
an attempt to reduce the problem of conflict of interest under uncertainty to a conflict of interest under risk
(Luce and Raiffa, 1957, p:14)
19
The vNM’s evaluation of utility can be thought of as a product of two factors: a measure of an increasing
or decreasing utilitarian marginal utility under certainty (labeled strength of preference) multiplied by an
intrinsic attitude towards risk. The relationship between the two is explored in Winterfeldt and Edwards
(1986).

84
probability function. Savage (1954) is a complete axiomatization of this earlier model of
expected utility for uncertainty and vNM (1944) is the complete axiomatization for risk.

This chapter presents first the vNM’s axiomatization of expected utility (EU) followed by
the more general representation of Savage’s (1954) subjective expected utility (SEU).
The axiomatization of expected utility under risk presented in this chapter is based on
Fishburn (1970), the literature on individual decision making under risk includes however
many axiomatizations (e.g. Herstein and Milnor, 1953; Jensen, 1967, Luce and Raiffa,
1957). Subjective expected utility by Anscombe and Aumann (1963) is sketched in brief.

1.2 EXPECTED UTILITY UNDER RISK

The axiomatization of expected utility in von Neumann and Morgenstern’s Theory of

Games and Economic Behavior (1944) is considered a pioneer work. The authors were
the first to axiomatize the model when the probabilities are given that is are objective and
their work is founded on the formalization of bets. The objective of this section is to
introduce the basic concepts and to present the axioms formally. It proceeds as follows:
first the general framework and some basic definitions are given, then the axioms and the
representation theorem are presented.

1.2.1 vNM Axiomatization of Expected Utility:

Axioms and Simple Probability Measures

vNM (1944) proposed a complete set of propositions and axioms necessary and sufficient
for the use of expected utility as a rule of choice under risk. In their work, the objects of
choice are probability distributions or lotteries defined on a given set of consequences.
The decision makers’ preferences are then formally represented by a binary relation.

The General Framework

Let X be the set of consequences and A a subset of X. A collection of subsets of X
denoted a is an algebra if X ∈ a; A ∈ a ⇒ − A ∈ a and A, B ∈ a ⇒ A ∪ B ∈ a (Kreps, 1988,
p: 116)

85
A probability measure P is a real-valued function which maps subsets of X into the
interval [0, 1] and satisfies the following axioms:

1- P (A) ≥ 0 for any subset A ⊂ X;

2- P(X) = 1;

3- P (A ∪ B) = P (A) + P (B) where A, B ⊂ X and A ∩ B = φ;

For a simple probability measure, a fourth property must be added:

4- P (A) = 1 for some finite subset A ⊂ X.

A simple probability measure P such as P({x1,…,xn}) = 1 for {x1,…,xn} ⊂ X, can be

considered as a lottery that gives xi with probability pi : P = (x1, p1,; …; xn, pn) for i =
1…n. Each consequence xi ∈ X can be represented by a degenerate probability measure
that is that gives the consequence with certainty.

If P and Q are simple probability measures on X and α ∈ [0, 1] then the convex
combination α P + (1 − α )Q can be interpreted as a combined lottery that gives the lottery
P with probability α and the lottery Q with probability (1- α ). It can be easily shown
that α P + (1 − α )Q is a simple probability measure on X. The set, P, of all simple
probability measures on X is closed under convex mixture operations, i.e. α P + (1 − α )Q
belongs to P is a mixture set.

The Axioms
The preferences of the individual over P the set of simple probability measures defined

over the set of the consequences X are represented by a binary relation ≽ that reads

“preferred to or equivalent to”. The binary relation ≽ is used to define two other binary

relations: The relation ∼ is called indifference and reads “indifferent or equivalent to”

i.e., P ∼ Q ⇔ P ≽ Q and Q ≽ P.

The relation ≻ is called a strict preference relation and reads “strictly preferred to” that
is, P ≻ Q ⇔ not Q ≽ P.

86
The binary relation ≽ is assumed to satisfy the following axioms for all P, Q, R ∈ P.

Axiom A1: Weak Ordering.

A1 is a fundamental tenet of rationality. It states that ≽ is a weak order on P, which

means that the relation ≽ is transitive i.e., P ≽Q and Q ≽ R ⇒ P ≽ R and complete i.e.,
the decision maker can compare lotteries in a consistent way such as P ≽ Q or Q ≽ P.
The relation ≻ is asymmetric by definition, and it can easily be established from A1 that
that the relation ∼ is transitive.

Axiom A2: The Independence Axiom.

P ≻ Q, 0 < α < 1 then α P + (1 − α ) R f α Q + (1 − α ) R

The independence axiom means that if Q is replaced by a preferred lottery P

in α Q + (1 − α ) R then the resulting compound lottery ( α P + (1 − α ) R ) should be preferred
to α Q + (1 − α ) R .

This axiom is also known as the substitutability axiom (Luce and Raiffa, 1957, p: 27). It
is strategic to the linearity in the probability property of expected utility and is often
considered as a principal normative criterion of the theory along with transitivity

(Fishburn, 1970).

Axiom A3: An Archimedean axiom.

P ≻ Q, Q ≻ R then α P + (1 − α ) R f Q and also Q f β P + (1 − β ) R for some α , β ∈

(0, 1).

This axiom rules out lexicographic preferences as well as unbounded utilities for
outcomes like heaven or hell. With A1 this axiom allows to establish the existence of an
ordinal utility function on X. It is important to the numerical representation of individual
preferences but is not normative as A2.

87
Theorem of Expected Utility under Risk

Theorem 1:
Suppose P is the set of all simple probability measures on X and ≽ a binary relation on P.
The following propositions are equivalent:

(i) ≽ satisfies the axioms A1, A2, and A3

(ii) There exists a real-valued function U which represents ≽ on P such that:

P ≽ Q ⇔ U(P) ≥ U(Q) ∀P, Q ∈ P (1.1)

U (α P + (1 − α )Q ) = αU ( P ) + (1 − α )U (Q ) , ∀α ∈ [ 0,1] , ∀P, Q ∈ P (1.2)

Moreover U is unique up to a positive linear transformation: that is if U on P satisfies

(1.1) and (1.2) then a real-valued function V on P satisfies (1.1) and (1.2) with U replaced

by V iff there are constants a > 0 and b such that V (x) = a U (x) + b ∀ P ∈ P.

Theorem 1 implies expected utility maximization on P. Let ∆ ⊂ P consist of all

degenerate probability measures i.e., P ∈ ∆ ↔ ∃ x ∈ X with P(x) = 1 and let elements of
∆ be denoted by δ x , x ∈ X. A function u on X can be defined from U on P the following

way. The preference relation is first extended to X:

δ x ≽ δ y ⇔ x ≽ y.

Then,

u(x) = U( δ x )∀ x ∈ X u being the restriction of U to X.

Hence, ∀ x, y ∈ X , x ≽y ⇔ u(x) ≽ u(y)

If ∀ P ∈ P, P(X) = 1, from the linearity property (equation 1.2) the following is

n
obtained: E (u, P) = ∑ p u( x ) .
i =1
i i

Hence, (1.1) and (1.2) ⇔ (1.3) the expected utility decision rule:

88
P≽ Q ⇔ E (u, P) ≥ E (u, Q) ∀P, Q ∈ P (1.3)

1.3 SUBJECTIVE EXPECTED UTILITY

Savage has succeeded in his Foundation of Statistics (1954) to treat uncertainty in an

entirely subjective manner. In comparison to vNM’s construction, there are more
conditions stated and some authors like Anscombe and Aumann (1963) have chosen an
intermediary road between vNM’s and Savage’s. This section proceeds as follows: first
Savage’s axioms and states of the world are presented. Next the phases in the
development of Savage’s theory are described followed by the formal statement of the
theorem of expected utility. A brief description of Anscombe and Aumann’s (1963)
theory is sketched subsequently.

1.3.1 Savage’s Axioms and States of the World

Savage’s (1954) subjective expected utility, a model of individual decision making under
uncertainty may be viewed as the result of combination of the expected utility of vNM
(1944) with its dual subjective probability theory (SP)20 of Bruno de Finetti (1937).
Savage (1954) constructed numerical subjective probabilities from comparative
subjective probabilities à la de Finetti (1937)21and showed that under his axioms, these
subjective probabilities obey the laws of probability.

20
SP where utility is linear and probability is subjective is the dual to SEU. Nau (2004) shows that the
axiomatization of SP in terms of acceptable p-gambles is equivalent to the axiomatization of EU in terms of
binary relations i.e. the axioms that the acceptable p-gambles must satisfy are the same vNM axioms as
reformulated by Jensen (1967, pp: 13).
21
For Bruno de Finetti, a degree of belief can be expressed vaguely as the extent to which an individual is
prepared to act on it. The individual’s degree of personal belief P in event E’s occurrence is revealed from
the amount he is willing to pay to play the lottery in which the payoff is W if E occurs. P = amount/W.
Once he bets, either there is the possibility that “he could have a book made against him by a cunning better
and would stand to lose in any event” (Ramsey, 1931) or there is no such possibility. In the latter case, the
individual’s evaluation of the probability is coherent. Otherwise it’s incoherent and presents an “intrinsic
contradiction.” The condition of coherence is the foundation for most of rational choice theory.

89
The Axioms
In Savage’s small world environment, the set of states of nature S is given and is an
exhaustive set of mutually exclusive states: one and only one state will be the true state of
the world but the decision maker is uncertain about which that will be. Events are subsets
of S and X is the set of outcomes or consequences. The objects of comparison are acts
which assign an outcome from X to each state of nature from S. They are denoted by f, g,
h ∈ A the set of all simple acts or finite-outcomes acts.22 The preference relation on A is

denoted by ≽ and reads “preferred to or equivalent to”. For whatever f ∈ A and B ⊂ S,

fB is an act that gives the consequence of f in B. For f and h ∈ B and 1-B, the object fBh is
the act which gives the consequences of f in B and the consequences of h in 1-B. For x ∈ X
the act xB f is the act f where all the consequences of the event B are replaced by x.
In addition, an event B is said to be null if any pair of acts which differ only on B are
indifferent. Savage’s theory of decision is deduced from the following axioms.23

P1. The relation ≽ represents a weak ordering on A i.e., transitive and complete.

This axiom is equivalent to the axiom A1 in vNM’s theory.

P2. Whatever f, g, h and h’ and the event B

fBh ≽ gBh ⇔ fBh’ ≽ gBh’

This is the sure thing principle, similar to A2 the independence axiom in vNM’s theory.

P3. For all non-null event B, the act f and the consequences x, y,
xBf ≽ yBf ⇔ x ≽ y

This axiom establishes a relation between preferences for acts and preferences for
consequences and is a natural companion to P2.

P4. For the consequences x, x’, y, y’ and if x ≻y and x’≻ y’ and B, C ⊂ S, then,

22
f(.) is said to be a finite-outcome act if its outcome set f(S) = {f(s)∣ s ∈ S } is finite (Machina and
Schmeidler, 1992).
23
Nau (2004) summarizes the axioms in a table which shows the similarity of the axioms of de Finetti
(1937), vNM (1944) and Savage (1954).

90
xB y ≽ xC y ⇔ x’B y’ ≽ x’C y’

Under this axiom, B is more probable than C. The probabilities are derived from choices
and the size of the outcome is irrelevant to the choice if the initial preferences do not
change.

P5. There is a least a pair of outcomes such that x ≻y.

This axiom ensures that the preference relation is not trivial.

P6. If f ≻ g and for all x ∈ X ⇒ there exists a finite partition of S such that if Bi and Bj
are any two events of this partition, then

xBif ≻ g and f ≻ xBig

If f is preferred to g there exists a partition that renders the act f constant for a
consequence x without varying otherwise, then the initial preference does not change.
This axiom establishes a link between subjective probabilities and objective probabilities
and gives to subjective probabilities the property of continuity essential to define a utility
function among acts.

Theorem of Expected Utility under Subjective Uncertainty

The first phase in the development of Savage’s theory is to obtain probabilities from
preferences. This is done in two steps. First, a qualitative probability is obtained.

Following Fishburn (1970), ≽p, a binary relation defined on 2s the set of all subsets of S
that reads “at least as probable as” is defined by:

B ≽p C ⇔ xB y ≽ xC y verifying x ≽ y; ∀ B, C ⊂ S

The second step is to show that the binary relation ≽p possesses under the Savage axioms
the attributes of a unique probability measure P* verifying:

(i) B ≽p C⇔ P*(B) ≥ P*(C) ∀B, C ⊂ S

91
 (ii) C ⊂ S, 0 ≤ λ ≤1 ⇒ P* (D) = λ P* (C) for a certain D ⊂ C

Having obtained subjective probability, the second phase is to show that the individual
attaches also a subjective utility to the acts’ consequences. P* is used by Savage to
construct a set of simple lotteries from the set of simple acts. The construction is based on
the idea that a simple act induces a simple probability measure on X. The simple lottery
induced by the simple act f ∈ A is denoted Pf and it is such that:

∀ x ∈ X , Pf (xi) = P*({s ∈ S: f(s) = xi }) = P * ( f −1 ( xi ))

Under the axioms P1-P6, it is shown24 that

Pf = Pg ⇒ f ∼g.
f and g are not the same, the implication is that f ∼x ∼g, thus f ∼ g. This condition

avoids having Pf = Pg and not (f ∼g).

The preference relation obtained allows to define a binary relation ≽ on the set of all
simple probability measures on X denoted P. For Pf and Pg from P, and f and g from A ,

the following binary relation is defined:

∀ f, g ∈ A : Pf ≽ Pg ⇔ f ≽ g.

The binary relation satisfies the vNM axioms25 and there exists a real-value function u
such that for Pf and Pg from P,

Pf ≽ Pg ⇔ ∑ u ( x) P ( x) ≥ ∑ u ( x)P ( x)
x∈ X
f
x∈ X
g

Or equivalently, for f and g from A

f≽g⇔ ∑ u ( x) P ( x) ≥ ∑ u ( x)P ( x)
x∈ X
f
x∈ X
g

24
Cf. Fishburn (1970)
25
Kreps (1988).

92
Formally, the above results can be stated in the following theorem.

Theorem 2:
Under the axioms P1- P6, there exists a subjective probability measure P* and a utility
function u defined on X and unique up to a positive linear transformation such as:

∀Pf , Pg ∈ P, f ≽ g ⇔ E (Pf, u) ≥ E (Pg, u)

Where
n
E ( Pf , u ) = ∑ u ( xi ) P * ( f −1 ( xi )) , {xi,…,xn} is the outcome set of the act f(.) and Pf (xi) =
i =1
−1
P * ( f ( xi )) = P*({s ∈ S: f(s) = xi }).

1.3.2 Anscombe and Aumann’s Axioms and Horse-Lotteries

Ten years later, Anscombe and Aumann (1963) presented another axiomatization of
expected utility under uncertainty. In their model, uncertainty is represented by a horse-
lottery rather than a Savage act. A horse-lottery is a mix of a vNM lottery and a Savage
act and is represented by lottery “akin in spirit” to the reduced compound lottery with the
distinction that one of the lotteries is a simple lottery where the probabilities are known,
as in a roulette wheel while the second is a horse lottery where the probabilities are
unknown, therefore personal as in a horse race (Anscombe and Aumann, 1963). The
order in which the two lotteries are run is immaterial because, Anscombe and Aumann
assume as Savage that the utilities of the consequences are the same in all the states. The
horse lottery corresponds to a Savage’s act, the outcome of the race to the state of the
world that obtains and the prize to the consequence; and the objects of comparison are
mappings from states to probability distributions over consequences.

93
In Retrospect

In the years that followed the axiomatization of expected utility, the normative axioms of
expected utility were recognized to be inadequate descriptively (Allais, 1953). The
research for the most part focused on the violation of the maximization of expected utility
and on the violation of the linearity in the probability property (Equation: 1.2). The
independence axiom A2 is implicit in both and its violation prompted theorists to
construct alternatives that either weakened the independence axiom or were “robust”
against its failures (Machina, 1982, p: 279). However, the two approaches which kept
unmodified the two other axioms A1 and A3 essential for the existence of a real-valued
utility function on X as in equation (1.1), could not explain some pervasive phenomena
most people exhibit most of the time (Kahneman and Tversky, 1979, p: 263). Prospect
theory (1992) was developed to accommodate four major phenomena of choice26 that
violate the standard model in addition to the violation of the independence axiom without
giving up the principle of coherence. Camerer (1997) shows its success in the field and
argues for its use along side the expected utility model in economic textbooks and in
current research, which confirms the important role the expected utility continues to play
despite its descriptive invalidity.

The resilience of the expected utility framework in the face of data challenging its
empirical validity is explained by Machina’s (1982, p: 277) characterization of the
vNM’s approach to the theory of individual behavior under risk: “the simplicity and
normative appeal of its axioms, the familiarity of the notions it employs (utility functions
and mathematical expectation) the elegance of its characterizations of different types of
behavior in terms of the properties of the utility function (risk aversion by concavity, the
degree of risk aversion by the Arrow-Pratt measure, etc.) and to the large number of
results it has produced.”

26
These are: the framing effect, the source dependence, the risk seeking and the loss aversion (Tversky
and Kahneman, 1992, p: 298)

94
The purpose of Chapter II the following chapter, is to provide an overview of the
violations of the properties of EU which are violations of its axioms and to set out the
alternative theories that can account for the descriptive invalidity of expected utility while
retaining the principle of coherence.

95
Chapter II: Violations of Certain Axioms of Expected Utility and the
Alternative Models

The experimental works which followed the theoretical works of vNM (1944) and
Savage (1954) showed that individuals violate systematically the axioms of expected
utility under risk and under uncertainty. The early violations were found by Allais (1953)
and Kahneman and Tversky (1979) for risk and by Ellsberg (1961) for uncertainty. In
these experiments the majority of individuals reversed their preferences. Once the
inconsistency was revealed, some decision makers revised their judgments to be
consistent with the axioms of the expected utility model but others refused to do so. In
the latter case, preference reversal spoke against the reasonableness of the related axiom
(Allais, 1953). The empirical evidence motivated researchers to develop alternative
theories of choice under risk and uncertainty to accommodate the observed violations.
Section 2.1 is concerned with the violations of the independence axiom. Section 2.1.1
presents the case for risk, and sections 2.1.2 and 2.1.3 present the cases for dynamic
context and uncertainty respectively. Section 2.2 presents the violations of the axioms
responsible for the stability of preferences and section 2.3 is concerned with the
alternative models developed.

2.1 THE VIOLATIONS OF THE AXIOMS OF EXPECTED UTILITY

The violation of the independence axiom under risk is the most discussed in literature and
also the one responsible for the generation of many alternatives to EU and SEU for a long
period of time stretching from 1979 till now. The reason is that without it the expectation
principle and the linearity in the probabilities (equation 1.2) cannot be retained. Section
2.1.1 focuses on its violation under risk followed by a brief description for what its
violation implies for dynamic contexts in section 2.1.2 and for uncertainty in section
2.1.3. Section 2.2 presents the violations of the axioms responsible for the stability of
preferences.

96
2.1.1 The Violation of the Independence Axiom under Risk

The independence axiom implies a separability property justified by the mutual

exclusiveness of the lottery’s consequences (Machina, 1989; Weber and Camerer, 1988).
The property is twofold: additive and multiplicative. The former implies that replacing a
common consequence with the same probability in two lotteries by a different
consequence does not influence the preference between the two lotteries. The latter
implies that multiplying all the probabilities in two lotteries by the same constant and the
remaining probability assigned to a common consequence, does not affect the preference
between the two lotteries. The violation of the separability property is known as the
common consequence effect and the common ratio effect respectively and in the
literature on dynamic choice it is known as the violation of consequentialism. The first
evidence of systematic violation of the separatibility property and hence the first evidence
of the inconsistency of individuals’ choices with equation (1.2) was found by M. Allais
(1952).

The Common Consequence Effect

The two sets of choices designed by Allais (1953) to demonstrate the violation of the
additive property of the independence axiom are described in Table 1 with M =
$1000000.

Table 1: The Allais Common Consequence Sets

P = 0.1 P = 0.89 P = 0.01

Set I
S1 1M 1M 1M
R1 5M 1M 0
Set II
S2 1M 0 1M
R2 5M 0 0

97
Many people presented with the above choices chose as predicted by Allais S1 over R1
attracted by the certainty of receiving 1 M and R2 over S2 because the consequences were
very different for quite similar probabilities. The choices however violate the
independence axiom. To see why, consider S1, R1, S2, R2 as lotteries and A and B as
intermediate lotteries such as A = (10/11) 5M + (1/11) 0 and B = 0, then

S1 = 0.11 S1 +0.89 S1, and R1 = 0.11 A +0.89 S1

S2 = 0.11 S1+ 0.89 B, and R2 = 0.11 A +0.89 B

According to the independence axiom, S1 ≻ R1 indicates that S1 ≻ A. However, the

preference R2 ≻ S2 indicates that A ≻ S1 which is a contradiction. Hence, the choices

observed by Allais are incompatible with the independence axiom.

Table 1 also shows why these choices are inconsistent with the behavior predicted by
expected utility. Going from Set II to Set I, all that is changed is that a 0.89 chance at
zero is replaced by a 0.89 chance at 1M, hence an expected utility maximizer would
choose S2 over R2 provided he has chosen S1 over R1 in the first choice otherwise there is
an inconsistency as shown by the following:

Assuming expected utility theory and u (5M) = 1 and u (0) = 0, S1 ≻ R1 yields: u (M) >
0.1 + 0.89 u (M) while R2 ≻ S2 yields: 0.1 > 0.11 u (M) which is inconsistent with u (M)
> 0.1 + 0.89 u (M).

The violation of expected utility was also observed in the absence of a certain
consequence as in Prelec’s (1990) example: with M = 10000, S1 = (2M, 0.02) and R1 =
(3M, 0.01); S2 = (2M, 0.34) and R2 = (3M, 0.01; 2M, 0.32). The common consequence
added here is (2M, 0.32). People chose S1 and R2 although the common consequence is
rationally irrelevant.

98
The Common Ratio Effect
Allais (1953) constructed also two other sets of choices to demonstrate the violation of
the multiplicative property of the independence axiom. The two sets are described in
Table 2.

Table 2: The Allais Common Ratio Sets

Outcome Probability
Set I
S1 3000 1
R1 4000 0.8
Set II
S2 3000 0.25
R2 4000 0.2

Confronted with Allais’ sets, many people chose S1 over R1 and R2 over S2 violating the
rule of constant probability ratio p/pq = α p/ α pq (in Allais’ sets, p = 1, q = 0.8 and α =
0.25) implied by the independence axiom27 and which justifies the choice of the same
lottery when the probability varies while the outcomes remain the same.

The certainty effect plays an important role in the reversal of choices; however even
when it is a special case of a more general form, people chose S1 over R1 and R2 over S2,
as in the following two sets of choices: Set I: S1 = ( x, p; 0,1 − p ) and R1 = ( y, q; 0,1 − q )
and Set II: S2 = ( x, α p; 0,1 − α p ) and R2 = ( y, α q; 0,1 − α q ) where 0 < x < y , p > q and
0 < α < 1, The certainty effect applies when p = 1.

The Transformation ofs the Probabilities

Prospect theory (1992) designed to explain the Allais paradoxes along with other
violations of EU incorporates a non-linear probability weighting of probabilities, p
→ w(p) also called a probability weighting function. Two properties required on this

27
( x , p )f ( y , pq ) implies that ( x , α pr ) f ( y , α pqr ) where 0 < p, q, r ≤ 1.

99
function reconcile the Allais preferences found inconsistent under EU: Subadditivity
explains the common consequence violation and subproportionality explains the common
ratio violation (Prelec, 2000).

Subadditivity
For the interpretation of the common consequence effect through probability weighting it
is convenient to describe the Allais (1953) sets in terms of decumulative probability
distributions. Therefore the set P of all simple probability measures on X defined in
section 1.2.1 is transformed into the set P* through a function (.)*: P → P* such that

(P)* = P* = ( p2* ,..., pn* ) , where pi* = ∑ j =i p j , i = 2,..., n. In other words, (.)* transforms
n

each lottery into a function P* assigning to each xi the probability of receiving xi or any
outcome rank ordered above in X (Abdellaoui, 2002).

Allais’ (1953) sets as alternatives in P and P* are described in Table 3. For x3 > x2 > x1,
P*, Q*, R* and S* are the decumulative distributions corresponding to P, Q, R and S
respectively. The table shows that the Allais preferences P*S* seem to indicate that
people tend to assign a greater weight to the replacement of probability 0.99 by
probability 1 than to the replacement of probability 0.1 by probability 0.11.

Table 3: The Allais Paradox (x1 = 0, x2 = 1M, x3 = 5M)

Alternatives in P Alternatives in P*
Set I P = (0,1,0) P* = (1, 0)
Q = (1/100, 89/100, 10/100) Q* = (99/100, 10/100)
Set II R = (89/100, 11/100, 0) R* = (11/100, 0)
S = (90/100, 0, 10/100) S* = (10/100, 10/100)

This particular interpretation is fostered by prospect theory (1992) where the Allais
preferences imply that CPT (P) > CPT (Q) and CPT (S) > CPT (R) and thus by definition

100
of CPT 28 that w(1) - w(0.99) > w(0.11) - w(0.1) that is the differential weight placed on
w(1) - w(0.99) is greater than the differential weight placed on w(0.11) - w(0.1). This
property of the weighting function is called subadditivity29 (Tversky and Wakker, 1995)
and it states that the increase in weight produced by adding probability ∆ to p is greater
when p + ∆ = 1 and certainty is reached than when p + ∆ < 1.

Subproportionality
Kahneman and Tversky’s (1979, p: 282)30 interpretation of the Allais pattern is that for a
fixed ratio, raising the chance of winning from 0.8 to 1 has a greater impact on relative
weight than raising the chance of winning from 0.2 to 0.25. Under CPT, the Allais
preferences imply: w (0.8)/w(1) < w(0.2)/w(0.25). This property of the weighting function
is called subproportionality and its preference conditions are a simple generalization of
the common ratio pattern observed by Allais (Prelec, 2000):

Subporportionality: for any 0 < α <1, y > x >0, and p > q, (x, p) ∼ (y, q) implies: (y,
α q) ≻ (x, α p).

The probability weighting function has one more empirical property and “perhaps the
first and most important empirical property” (Prelec, 2000) which says that small
probabilities are overweighted and large probabilities are underweighted.

28 +
CPT’s valuation of ( x , p ) = w ( p )u ( x ) for x > 0. Chapter III is devoted to CPT.
29
Subadditivity also states that the increase in weight produced by adding ∆ to p is greater when p = 0 than
when p > 0. The formal definition is given in section 3.1.3.
30
The violation of the multiplicative property of the independence axiom conforms to the following rule: if
(x, p) is equivalent to (y, pq) then (x, pr) is not preferred to (y, pqr), 0 < p, q, r ≤1. By definition of the
prospect theory utility function i.e. by equation (3.1):
π (p) u(x) = π (pq) u(y) implies π (pr)u(x) ≤ π (pqr)u(y); hence,
+
( x , p ) = w ( p )u ( x )
π (pq)/ π (p) ≤ π (pqr)/ π (pr)

101
Along with the unattractiveness of probabilistic insurance,31 it has led Tversky and
Kahneman (1992) to the four-fold pattern of risk attitudes:

“Overweighting of small probabilities contributes both to risk seeking for gains and risk
aversion for losses when the outcomes are high which explains gambling (optimism) and
insurance (pessimism). Underweighting of the probabilities contributes to the prevalence
of risk aversion in choices between probable gains and sure things and to risk seeking in
choices between probable losses and sure losses.”

The fourfold pattern is empirically supported by later parametric studies: Tversky and
Kahneman (1992); Camerer and Ho (1994); Tversky and Fox (1995); Wu and Gonzalez
(1996); Abdellaoui (2000) and non-parametrically, Abdellaoui (2000) and Bleichrodt and
Pinto (2000).

2.1.2 Implication of the Violation of Independence for Dynamic contexts

The independence axiom is also assumed in principles of dynamic choice. Wakker (1999)
shows that four conditions jointly imply equivalence of independence in dynamic
contexts. It follows in view of the evidence for the failure of independence that at least
one principle of dynamic choice must be failing too. To recall from section 1.2.1 the
independence axiom A2 states:

P ≻ Q, 0 < α < 1 then α P + (1 − α ) R f α Q + (1 − α ) R

Figure 1 illustrates the independence axiom. Squares denote decision nodes, circles
chance nodes and arrows the preferred path.

31 p p
That is reducing a probability from p to is less valuable than reducing to zero. Allais’ gamble also
2 2
illustrates that individuals are oversensitive to changes in small probabilities.

102
Figure 1: Independence

P αP+(1-α)R
Independence

Q αQ+(1-α)R

Because the verbal expressions of the conditions are often ambiguous32, the decision
making process in Figure 1 is separated as in Wakker’s (1999) into four stages to
illustrate each condition.

The first condition is forgone-event independence referred to as consequentialism by

Machina (1989). It is helpful to draw at this point the choice between P and Q as in
Figure 1b. Indeed, the latter illustrates the same choice as is 1a because the options and
their consequences are the same.

P
P α
Forgone-Event
Independence
Q

Q 1-α
R
Fig. 1a Fig. 1b

Forgone-event independence states that if one arrives at the decision node in Figure 1b
the lower branch is irrelevant to the consequences of choices at that decision node.

32
The terminologies also vary: Hammond (1986) uses the term consequentialism (currently mostly used for
foregone-event independence, Machina, 1989) for equivalences of Figures 1b, 1c and 1d with 1e. Burks
(1977) calls the equivalence of Figures 1a and 1c invariance and the equivalence of figures 1c and 1d with
1e normal-form equivalence (Wakker,1999)

103
The second condition is dynamic consistency and is illustrated in Figures 1b and 1c. It
requires that the decision maker commits before the resolution of uncertainty (hence the
dashed lines in Figure 1c) at the chance node to a decision at a future decision node and
does not deviate from his prior commitment once he reaches the decision node.

P P
α α
Dynamic
Consistency
Q Commit to Q
decision
already here

1-α 1-α
R R
Fig. 1b Fig. 1c

Context–independence is the third condition and is illustrated in Figures 1c and 1d. In the
latter, the upper branch depicts the prior commitment of going up and the lower branch
depicts the prior commitment of going down. The two figures depict the same situation
and hence should be treated the same.

P α P
α
Context
Independence 1−α R
Commit to Q
decision
already here
α Q
1-α
R
Fig. 1c Fig. 1d 1−α R

The last condition is illustrated in Figures 1d and 1e. Figure 1e depicts each lottery at the
two chance nodes in Figure 1d as probability distributions over consequences. Hence,
the two figures should also be treated the same. This condition is called reduction.

104
 α P
αP+(1-α)R
Reduction
1−α R

α Q
αQ+(1-α)R
Fig. 1d 1−α R Fig. 1e

The four conditions together imply the equivalence of Figures 1a and 1e, i.e.,
independence.

Among the studies that have addressed the issue of which condition must be relaxed in
order to predict actual behavior are Machina’s (1989) and McClennen’s (1990) who
relaxed consequentialism and Segal’s (1990) who relaxed reduction.

2.1.3 Violation under Subjective Uncertainty

The violation of the independence axiom under uncertainty which is a violation of

Savage’s axioms P2, the sure thing principle33 was first shown by Ellsberg (1961) who
designed the following two sets of acts described in Table 4.

Table 4: The Ellsberg Paradox

States of Nature
Red Black Yellow
30 60
Set I
Act I 100 0 0
Act II 0 100 0
Set II
Act III 100 0 100
Act IV 0 100 100

33
P2 with P3 constitute the independence axiom in Savage’s axiomatization of expected utility.

105
Subjects are presented with an urn known to contain 30 red balls and 60 black and yellow
balls in unknown proportions and asked to choose first, an action from Set I, then an
action from Set II.

The acts of Set I and the acts of Set II have a common consequence as in Allais’ example
and differ from the latter only with respect to the information regarding the states of
nature of the black and yellow balls. The latter are considered ambiguous because the
information about either state of nature is unknown.

The most frequent pattern, Act I (preferring to bet on red) and Act IV (preferring to bet
against red) violate the sure thing principle. Indeed, Act I implies according to expected
utility under uncertainty and after simplification that π ( R ) > π ( B ) and Act IV implies
that π ( B ∪ Y ) > π ( R ∪ Y ) . Because π is additive, the implication of Act IV is that
π ( R ) < π ( B ) which is a contradiction with the implication of Act I.

Ellsberg’s (1961) explanation for the choices and the contradiction is that people’s
“unease with their best estimates of probabilities” make them prefer objective rather than
subjective probabilities which conflicts with P2’s implication that once people form their
subjective probabilities they use them exactly as objective probabilities. Aversion for
ambiguity (absence of information about the states of nature of black and yellow) which
refers to the above “unease” is well documented in Heath and Tversky (1991), Camerer
and Weber (1992) and Tversky and Fox (1995).

2.2 VIOLATIONS OF THE AXIOMS OF STABILITY OF PREFERENCES

The axioms A1 and A3 are essential for the existence of a continuous real-valued
preference function. The Archimedean axiom is not normative, however relaxing it leads
to lexicographic preferences implying extreme attributes and restricted situations.
Regarding completeness some authors Nau (2004) favor “dropping” it to relax the
requirements on transitivity and independence two normative criteria. Transitivity

106
ensures that the decision maker is not transformed into a money pump,34 however, it has
been found to be violated as well as two other normative criteria implicit in the axiom of
ordering: procedure invariance and description invariance.

2.2.1 Transitivity
The earliest evidence on intransitivity goes back to Georgescu-Roegen (1936) and May
(1954). Tversky (1969) and Tversky, Slovic, and Kahneman (1990) also show that
cyclical choice is robust. Regret Theory (Bell, 1982; Fishburn, 1982; Loomes and
Sugden, 1983) is an attempt to remedy to intransitive behavior, but there is no theory yet
that is fully consistent with the available data.

2.2.2 Procedure Invariance

Procedure invariance says that preferences over lotteries are independent of the method
used that is, it implies that people follow the same procedure for valuing and choosing.
Evidence of its violation was provided by Lichtenstein and Slovic (1971), Harold
Lindman (1971) and Grether and Plott (1979). Individuals, repeatedly, chose probability
bets that offer a larger probability of winning a smaller prize while valuing more dollar
bets that offer a small probability of winning a good prize. Tversky, Slovic and
Kahneman (1990) show that the preference reversal is due to procedure invariance
violation rather than to the violation of either of transitivity or the independence axiom or
the violation of the reduction of lottery axiom. The violation of the procedure invariance
violation implies the existence of two different processes for valuing and choosing and is
supported by Slovic and Lichtenstein (1983) who provide evidence that the latter is
fundamentally influenced by the probabilities associated to the outcomes while the
former is fundamentally influenced by the outcomes.

34
Fishburn (1970, p: 108-109) and Machina (1989) illustrate the normative appeals of transitivity and the
independence axiom. Transitivity also ensures that indifference curves are reversible and do not cross.
Knetsch (1990) however demonstrates that owning A and being indifferent about trading it for B is not the
same as owning B and being indifferent about trading it for A in the presence of loss aversion. In that case,
the indifference curves intersect (Kahneman, Knetsch and Thaler, 1991).

107
2.2.3 Description Invariance
According to Starmer (2000) description invariance is an implicit assumption in any
conventional theory and seems so natural to most economists that it is rarely discussed.
Descriptive invariance says that preferences over lotteries are purely a function of the
probability distributions of consequences implied by the lotteries and do not depend on
how these given distributions are described. Nevertheless, the framing effect provides
evidence that this assumption fails in practice.

The Framing Effect

Ramsey (1931) referred to the possibility that choice depends on the special form of the
offered options as “absurd”. However, Kahneman and Tversky’s (1979) Asian disease
shows that people are influenced by the way questions are framed: the choices are
reversed when outcomes initially framed in terms of lives saved are framed in terms of
lives lost. The preference reversal implies that the assumption of descriptive invariance is
violated in response to what has become to be known as a framing effect. Evidence of
framing effects is also cited by Slovic (1969), Schoemaker and Kunreuther (1979),
Tversky and Kahneman (1981, 1986) and in (Popkin, 1992).

2.2.4 The Reflection Effect

Although the reflection effect is not a violation of description invariance because it
involves different options, it is consistent with the framing effect: Kahneman and
Tversky (1979) found in one of their experiments that when the sign of outcomes in
problems involving positive prospects is reversed, responses change to the exact
opposite, that is, behavior towards losses becomes the mirror image of behavior towards
gains. Indeed, given the Allais (1953) example described on page 99 with the signs
reversed, Kahneman and Tversky’s (1979, p: 268) participants reversed their responses.
For instance, those who prefered sure gains to probable gains preferred also probable
losses to sure losses. The authors labeled this reversal of preferences around 0 the
reflection effect. Budescu and Weiss (1987) found 82% of their subjects displayed
concavity for gains and convexity for losses. Related evidence include: Fiegenbaum and
Thomas (1988); Lowenstein (1988), Terence Odean (1988); Platt and Glimcher (1999);

108
Smith et al. (2002); Breiter et al., (2001); and Abdellaoui et al. (2006). Yet, the evidence
is not indubitable; many empirical studies provide evidence for a linear utility for losses
(Edwards, 1955; Hershey and Shoemaker, 1980; Schneider and Lopes, 1986; Cohen,
Jaffray, and Said, 1987; Weber and Bottom, 1989; and Lopes and Oden, 1999).
Moreover, Levy and Levy (2002) argue for a utility function that is convex for gains and
concave for losses.

Sections 2.1 and 2.2 reviewed several empirical effects which seem to invalidate
expected utility as a descriptive model. According to Starmer (2000, p: 332) the number
of alternative models stimulated by these violations is well into double figures for, to
many economists “put bluntly, the standard theory did not fit the facts”.

2.3 THE ALTERNATIVE MODELS

In view of the above violations, many economists concluded that expected utility theory
either does not correspond to the facts at all or correspond to only some of the facts,
despite arguments proceeding with faultless logic from the postulates to the conclusion.
For the development of a serious contender at least for some purposes, decision theorists
revisited and revised the axioms of EU and doing so generated a large number of theories
going hand-in-hand with ongoing experimental tests of these theories. Section 2.2.1
presents an overview of the recent developments in utility theory and section 2.2.2
presents the alternative models.

2.3.1 Overview of Recent Developments in Utility Theory

This section builds on extent overviews of these alternative theories and their
experimental tests (Camerer, 1989; Schmidt, 2002; and Starmer, 2000) and focuses on
the models which according to these summaries account best for the currently available
empirical data.

109
Thus, many important alternatives will not be considered. For instance, theories with the
betweenness property will not be considered because of the empirical evidence against
linear indifference curves implied by the property35. Machina’s (1982) generalized
expected utility will not be considered also because the generalized fanning-out36 implied
by his hypothesis II is ruled out by the presence of fanning-in in numerous studies37. The
alternative theories presented in this section are thus restricted to the rank dependent
weighting models (RDU, 1982; CPT, 1992) which allow mixed fanning and are not linear
in the probability,38 specially in view of the striking degree of convergence across studies
regarding the form of their probability weighting functions (Starmer, 2000, p: 359).

To emphasize however, the inspirations and ideas the authors drew from each other and
the intuitive leaps in convergent thinking promoted by combining extent knowledge and
/or tolerating dualities39 prospect theory (1979), the original form of prospect theory
(1992) is reviewed first. Quiggin’s (1982) anticipated theory, developed in part to build
some of the non-conventional prospect theory’s features along a conventional strategy is
presented next. Yaari’s (1987) dual theory (DT) a special case of Quiggin’s which is
linear in consequences and non linear in the probabilities and therefore the dual of EU,
follows. Prospect theory (1992) a cross fertilization of prospect theory (1979) and
anticipated theory (1982) is then contrasted with both. However, prospect theory (1992)’s
psychological elements, the formal model and its axiomatization for risk are presented in
Chapter III in view of their pertinence to the experimental part of this work. Chapter III’s
also presents the theories’ respective characterizations for risk aversion.

35
Camerer and Teck-Hua Ho (1994) provide evidence against the linearity of indifference curves.
36
Fanning out means the indifference curves become steeper (bigger slope) as one moves northwest in the
probability triangle, as opposed to fanning in which means the indifference curves become flatter (the slope
becomes smaller) in the right hand side of the probability triangle.
37
Camerer (1989), Chew and William Waller (1986) and Starmer (1992) provide evidence for fanning in.
38
Table 6 in the Appendix based on Starmer (2000) compares Machina’s indifference curves with those
implied by theories with the betweeness property and rank dependent theories.
39
A dual is a connection between two problems which turn out be the very same problem looked at from
different angles; for instance Yaari’s (1987) theory is the dual of vNM’ theory (1944). Duality is borrowed
from physical chemistry. Light is dual in the sense that in some experiments its wave properties are most
obvious, in others it behaves as a particle.

110
2.3.2. Theoretical Predictions

The main concern being the reconciling of the theories’ predictions with the experimental
facts, the presentation is narrowed to a comparison of the forms of the indifference curves
between sets of gambles under the different theories and those conjectured from
observations in the lab and/or in the field. The indifference curves under expected utility
are discussed first to be contrasted with those generated from choices that violate its
independence axiom. The indifference curves under the alternative theories developed in
response to the violations follow. All indifference curves are presented in the probability
triangle in order to compare them visually.

The Probability Triangle

When a lottery L = (x1 p1; x2 p2; x3 p3) with fixed outcomes such that x1> x2 > x3 is
considered in a probability triangle where to recall, the horizontal side represents the
probability of the worst consequence x3 and the vertical side the probability of the best
consequence x1 and where the probability of the third consequence is deduced: p2 = 1 –
(p1+ p3), the set of all possible lotteries is contained in the triangle with the vertical side
(left edge) characterized by a zero probability for the worst consequence, the horizontal
side (lower edge) by a zero probability for the best consequence and the hypotenuse by a
zero probability for the middle consequence.

An indifference curve is the set of lotteries with the same utility. Therefore,
differentiating totally the utility function with respect to p1 and p3, and setting the
derivative equal to zero to maintain the utility constant gives the set of indifference
dp1
curves that characterize the preference function. The ratio is the slope of the tangent
dp3
line to the indifference curve and is an indicator of its shape at any point in terms of the
components of the numerator and the denominator.

111
Expected Utility

Axioms A1 and A3 of expected utility theory imply that preference functions can be
represented by well-defined indifference curves in the space of the probability triangle.
A1 implies that lotteries either lie on the same indifference curve or on different
indifference curves (completeness) and that these do not cross inside the triangle
(transitivity) and A3 implies that these are not thick and there are no holes in the
indifference curves map. A1 and A3 do not impose however, any restrictions on the form
of the indifference curves.40

The utility function of a lottery under expected utility is computed according to:

n
EU ( L) = ∑ p j u ( x j )
j =1

Accordingly, the utility of lottery (L) is:

EU (L) = p1 u (x1) + p2 u (x2) + p3 u (x3) (2.1)

dp1 u ( x2 ) − u ( x3 )
= (2.2)
dp3 u ( x1 ) − u ( x2 )

The slope is constant because it’s independent of p and implies the indifference curves
are linear and parallel with northwest movements along the hypotenuse leading to
increasing preferences, i.e. to a higher utility level. The independence axiom A3 implicit
in the linearity property of expected utility theory restricts the indifference curves to
linearity, parallelism, and upward sloping leaving however the slope undetermined.

Figure 2 shows the indifference curves under EU and also shows the Allais’ lotteries [S1,
R1, S2, R2] described in Table 1 above. It can be easily shown that the two lines which
join the pairs of lotteries involved in the two choices are parallel. EU allows the

40
Positive affine transformations v(x) = a u(x) + b, a > 0, represent preferences equivalently because the
origin and the scale can be arbitrarily defined.

112
indifference curves to exactly coincide with the lines joining the lotteries [S1 ∼R1 and S2,

∼R2], to be less steep [S1 p R1 and S2 p R2] or steeper [S1 ≻R1 and S2 ≻R2].

Figure 2: Expected Utility Indifference Curves

s2>r2
r1 r2
1

s1>r1
s1 s2

However, the indifference curves generated by actual preferences of the Allais’

experiments that is, S1 ≻ R1 and R2 ≻ S2 are not parallel. They are flatter in the right hand

corner (S2, R2) relative to those in the left hand edge of the triangle (S1, R1). In other
words, the preference in the right hand corner is contrary to the prediction of EU given
the preference near the left hand border. To conform to actual behavior, indifference
curves in any contender for expected utility need to be flatter in the right hand corner
like for instance relative to those in the left hand edge of the triangle like.

Original Prospect theory

Original Prospect theory (Kahneman and Tversky, 1979) is a theory designed to account
for psychological insights the authors found pervading laboratory and field data (for

113
instance probability transformation) and which are not accounted for in the normative
EU. The authors’ aim is not as much to find out whether or not individuals are rational
but to provide a descriptive model for actual individual behavior under risk. This review
of prospect theory (1979) is based on Camerer (1989)41 which shows an illustrative figure
of prospect theory’s indifference curves.

Original Prospect theory (PT) differs from expected utility in four points: 1) It applies
only to lotteries, referred to as prospects by the authors, with at most two non zero
outcomes; 2) the outcomes are “coded” that is perceived as gains or losses relative to a
reference point and not final assets with losses looming larger than corresponding gains;
3) lotteries are edited to make them simpler to evaluate for instance using the rule of
combination or cancellation for common outcomes; and 4) edited lotteries are evaluated
according to one of several expectation like rules that combine the u(x), or the v(x) in the
authors terminology,42 and a decision weight π(p) which transforms the probability non-
linearly. The decision weight is increasing, subadditive, i.e., a change in probability has
less impact as one moves away from the boundaries to the middle, (π (p) + π (1-p) < 1)
and discontinuous at the end points 0 and 1. The utility function defined on i.e. deviations
from the reference point is generally concave for gains and convex for losses and steeper
for losses than for gains.

If x3 = 0 and outcomes x2 and x1 are either both gains x1 > x2 > 0 or both losses 0 > x2 > x1
relative to the reference point and p2 + p1 < 1 (p3 > 0), the edited lottery is evaluated
according to:

PT (L) = π (p2) u (x2) + π (p1) u (x1) (2.3)

Camerer (1989, p: 75) shows that the slope is equal to:

dp1 π '( p2 )u ( x2 )
= (2.4)
dp3 π '( p1 )u ( x1 ) − π '( p2 )u ( x2 ))

41
Camerer (1989, footnote 10) explains how prospect theory’s indifference curves were computed.
42
The utility terminology is used throughout this work; that is there is no reversion to the term “value” to
describe utility under prospect theory.

114
The slope expressed in terms of the decision weight function and of the utility function
varies according to the value of p in the numerator and in the denominator causing the
indifference curves to be steeper (fan out) in some regions of the triangle (for instance in
the lower left hand corner where p2 = 1) and flatter (fan in) in others (for instance near the
hypotenuse where p2 is 0 and fixed).

The indifference curves appear on the hypotenuse and near the lower edge very steep and
flat respectively. The pattern reflects the preference people have for points inside those
edges to points exactly on the edge which in turn reflects people’s overweighting of the
small probabilities of winning x2 near the hypotenuse or x1 near the lower edge.

Thus, indifference curves under prospect theory are not linear (π (p) is assumed to be
convex by Kahneman and Tversky except near the end points), do not uniformly fan out
and reflect sensitivity to extreme probabilities; however, they allow violation of
monotonicity or first order stochastic dominance,43 the most widely acknowledged
criterion for rationality: In the lower left-hand corner, the negatively sloped part of the
indifference curve shows that some lotteries stochastically dominate others but all are
equally preferred. Kahneman and Tversky assume that people eliminate in the editing
phase dominated lotteries if the dominance relation is transparent. The “if” however,
leaves open the possibility of intransitive choice among three lotteries. Figure 3 shows
the indifference curves of prospect theory (1979).

43
∀ F, G ∈ D(X) the set of all cumulative probability distributions functions over X, a lottery F is defined
to dominate a lottery G by first order stochastic dominance (F ≥1 G) if F (x) ≤ G (x) ∀x ∈ X and F (x) <
G(x) for at least one x ∈ X.
The definition in terms of cumulative distributions is used for consistency with later sections. There is a
one-to-one correspondence between the set P of all probability measures and the set D(X) (Herstein and
Milnor, 1953). Put differently, F ≥1 G if the probability that any x is less than xi under G is greater or equal
than the probability that any x is less than xi under F with at least a strict inequality.

115
 Figure 3: Indifference Curves Assuming Prospect Theory (1979)

Rank Dependent Utility

Quiggin’s (1982) anticipated utility theory is the first model that incorporates a
probability weighting function and a utility function without violating monotonicity as in
(Handa, 1977; Kamarkar, 1979; Kahneman and Tversky, 1979). According to Quiggin,
the fundamental problem in these theories is that any two outcomes with the same
probability need not have the same decision weight; hence, in cases where extremes
outcomes are overweighted, at least some intermediate outcomes perhaps with the same
objective probability must be underweighted (Quiggin, 1982, p: 326-328). To formalize
this insight, the entire cumulative distribution was transformed and each outcome
weighted according to its rank relative to other outcomes by a discrete chunk44 of the
transformed cumulative distribution, ensuring monotonicity. The outcomes in each

lottery are ranked ordered such as x1 ≥ x2 ≥…≥ xn and each outcome is weighted by a

44
Or by differentials if the cumulative distribution function is continuous.

116
decision weight depending not only on the probability of the outcome but also on its
rank, hence the name of rank dependent utility theory (RDU).

The utility of a lottery according to RDU is given by:

n
RDU ( L) = ∑ π j u ( x j ) (2.5)
j =1

Where

j j −1
π j = w(∑ pi ) − w(∑ pi )
i =1 i =1

For all j.

The decision weights π j sum to 1 and depend on the ranking of the outcomes. The

function w (.) is the decision weight generated by the probability p when associated with
the best outcome such that,

For each j,

π j = w (p1 +…pj)- w(p1+…pj-1) with π 1 = w+ ( p1 ) for j = 1.

One could choose however to use w* (p) the dual of w (p) that is w* (p) = 1- w (1-p) for
all p, which is the decision weight generated by the probability p when associated with
worst outcome such that,

For each j,

π j = w* (pj +…pn)- w*(pj+1+…pn) with π n = w * ( pn ) for j = n.

This duality follows because the decision weights sum to one for any lottery (M, p; m, 1-
p) with outcomes M > m. w can be called the goodnews weighting function and w* the
badnews weighting function (Diecidue and Wakker, 2001).

117
The probability weighting function w(.) is a strictly increasing function from [0, 1] to
[0, 1] and verifies w(0) = 0 and w(1) =1. The weighting function has the shape of an
inverted S with, w(0.5) = 0.5, w( p ) > p if p < 0.5, w( p ) < p if p > 0.5, Quiggin (1982)
having situated the cross over at 0.5.

The lottery L is evaluated according to the following equation:

RDU (L) = w( p3 )u ( x3 ) + [ w( p3 + p2 ) − w( p3 )] u ( x2 ) + (1 − w( p3 + p2 ))u ( x1 ) (2.6)

Camerer (1989, p: 77) shows that the slope is equal to:

dp1 w '( p3 )(u ( x2 ) − u ( x3 ))

= (2.7)
dp3 w '(1 − p1 )(u ( x1 ) − u ( x2 )

and Figure 4 shows that the indifference curves (the weighting function is assumed
convex) are steepest in the left corner (fan out) and get flatter (fan in) as one moves along
the lower edge p1 = 0 or the left edge p3 = 0 (vertically upwards) The curves are equal in
slope along p2 = 0 this property is called hypotenuse parallelism. Thus, the curves do not
uniformly fan out, are not linear, reflect sensitivity to extreme probabilities and do not
violate monotonicity.

118
 Figure 4: Indifference Curves Assuming RDU

Expected utility is the special case of RDU where there is no probability weighting,
w( p ) = p for all p ∈ (0, 1) and equation (2.7) is reduced to equation (2.2).

Dual theory (Yaari, 1987) is a special case of anticipated utility theory where u(x) = x.
Yaari’s theory has been developed independently of Quiggin’s and is the object of the
nest section.45

45
Green and Jullien (1988) and Segal (1989) have axiomatized independently a general model of rank
dependent utility which contains anticipated utility and dual theory as special cases.

119
Dual Theory

Dual theory is so called because it’s dual to expected utility theory. In contrast to the
latter, it’s linear in the utility and non-linear in the probability. This duality implies that
dual theory is deduced from the same axioms of EU with the sole difference that the
independence axiom is “laid on its side” (Yaari, 1987, p: 98). Rather than being assumed
over convex combinations of probability measures and implying linearity in the
probabilities, it is assumed over convex combinations of consequences and implies
linearity in the consequences. The following equations compare the functional forms of
dual theory, expected utility and rank dependent theory for ranked outcomes such as

(x1 ≥ x2 ≥…≥ xn):

n
DT(L) = ∑π j =1
j xj (2.8)

n
EU(L) = ∑ p u( x )
j =1
j j

n
RDU(L)= ∑ π u( x )
j =1
j j

j j −1
With π j = w(∑ pi ) − w(∑ pi ) .
i =1 i =1

As seen from the above equations, dual theory with u (xi) = xi, has the merit of isolating
the implications of the weighting function for risk aversion which is one of the two
reasons that prompted Yaari (ibid, p: 95) to look for an alternative to expected utility.46
Concerning EU’s empirical violations, Yaari’s second motivation, DT accommodates the
common consequence effect and the common ratio effect, but as Yaari (ibid, p: 96)
acknowledges, behavior inconsistent with the linearity of the utility is often observed.
Yaari (ibid, p: 108) also emphasizes that DT (as well as RDU for that matter) deals with
how perceived risk is processed into choice and not how actual risk is processed into
perceived risk and thus is not concerned with the violations due to perceptual causes.

46
The characterization of risk aversion under the different theories is discussed in chapter III.

120
RDU, as is its special case DT, is thus focused on the mathematical connections between
the axioms and the numerical representations of preferences and is less descriptive than
the non-conventional model PT (1979) characterized by a procedural approach and
reference dependence, an approach more common to psychology than to economics. The
case for both suggested by Starmer (2000) seems to plead as in Camerer and Weber
(1992) for a communication between psychologists and decision theorists whereby the
former benefit from the latter in mathematical precision and the latter from the former in
descriptive validity. Cumulative prospect theory (1992) a cross fertilization of the non-
conventional prospect theory (Kahneman and Tversky, 1979) and the conventional
anticipated utility theory (Quiggin, 1982) establishes such a common language.

Prospect Theory (1992)

Quiggin’s (1982) RDU showed that decision weights constructed cumulatively eliminate
the possibility of intransitive choice among three lotteries left open by PT (1979). By
incorporating the idea, Kahneman and Tversky were able to obtain a transitive and
monotonic preference function that generalizes to n-outcomes prospects without
assuming the editing phase (Tversky and Kahneman, 1992).

Prospect theory (1992) differs from prospect theory (1979) in four points: 1) it applies to
prospects with an arbitrary number of outcomes rather than to two non-zero outcomes at
most; 2) although the outcomes are still “coded” that is perceived as gains or losses
relative to a reference point with losses looming larger than corresponding gains, no
editing phase is required; 3) the utility function in the evaluation rule is the same as in PT
but the decision weight function transforms the entire cumulative distribution and needs
not be the same for gains and losses; 4) CPT applies to both risk and uncertainty. Hence,
the main difference between PT and CPT is a transformation of cumulative probabilities
rather than of individual probabilities, which makes CPT a rank dependent theory since
the decision weight attached to an outcome depends on the rank of the outcome relative
to the other outcomes.

121
Nevertheless, it is a more descriptive and a more general representation. A comparison
with anticipated utility shows that CPT differs from the latter in two respects: 1) the
utility function is defined on changes in wealth (gains and losses) rather than on final
wealth with losses looming larger than corresponding gains; 2) there are two weighting
functions, one for gains and one for losses and the transformation is on decumulative
probabilities in the gain domain and on cumulative probabilities in the domain of losses.
Hence, the weighing function under CPT is equal to the sum of two weighting functions
under anticipated utility which are computed separately for gains and losses. RDU
corresponds to the special case where the weighting function for the losses is the dual of
the weighing function for gains: w− ( p ) = 1 − w+ ( p ) . The shape however of the weighting
function is the same inverse S, under the two models.

CPT’s indifference curves shown in Figures 5(a), 5(b) are constructed by Tversky and
Kahneman (1992, p: 314). They are non linear, present mixed fanning and do not violate
monotonicity. The curves for positive lotteries resemble those for negative lotteries but
are not the same.

Figure 5 (a): Indifference Curves Assuming CPT for Positive Lotteries

122
 Figure 5 (b): Indifference Curves Assuming CPT for Negative Lotteries

Two important features distinguish thus CPT from RDU: 1) the utility function defined
relative to a reference point is characterized by non-normative properties pertaining to
perception and judgment as in losses looming larger than gains; 2) CPT generalizes rank
dependent utility by allowing for different risk attitudes. Chapter III presents first the
descriptive model followed by the formal model and its axiomatization for risk. The
characterization of risk attitude in the different models is consigned to the last section of
chapter III.

123
CHAPTER III. CUMULATIVE PROSPECT THEORY

“The goal that we set for ourselves was to assemble the minimal set of
modifications of expected utility theory that would provide a descriptive account
of everything we knew about a severely restricted class of decisions: choices
between gambles.”

(Kahneman and Tversky, 2000, preface: x)

What they knew from observations and a series of experimental and empirical
investigations was that choices between gambles presented patterns that could be related
to a simple perceptual cause: that of a reference point relative to which outcomes and/or
probabilities are considered. They suggested new hypotheses which they were able to
verify and provided labels for the patterns to ease their identification in more complex
contexts. Two of these phenomena: the reflection effect and loss aversion pertain to the
utility function and underlie the empirical realism relative to which prospect theory is
considered a better description of individuals preferences than rank dependent theory.
Although, this work is concerned primarily with the elicitation of the utility function, the
weighting function is described in order to emphasize the importance of filtering its
impact on the desirability of the outcomes. Section 3.1 presents the descriptive model
followed by the formal model in section 3.2. Finally, the axiomatization for cumulative
prospect theory for risk is presented in section 3.3 based on Chateauneuf and Wakker
(1999).

3.1 THE DESCRIPTIF MODEL

Two principles, diminishing sensitivity to deviations from a reference point and loss
aversion are invoked by Tversky and Kahneman (1992) to explain the characteristics of
the utility function and the weighting function of cumulative prospect theory. The
objective of this section is to present these non-normative elements in the theory. It

124
proceeds as follows: first the intuition for the reference point is given and the importance
of its localization for the order of preferences is emphasized. Loss aversion is presented
next followed by the implications of these non-normative elements for the utility function
and the weighting function in sections 3.1.2 and 3.2.3 respectively.

3.1.1 THE NON-NORMATIVE ELEMENTS IN THE THEORY

The Reference Point

The intuition of the reference point is best understood in relation to psychophysics.
Psychophysics is the mapping of physical stimuli into psychological responses (Stevens,
1957; Sinn, H.1983)47 and is characterized by diminishing sensitivity: discriminability is
good in the central range where the most frequent stimuli occur and the mapping is
almost linear; however, at the very high (low) stimuli, the sensitivity diminishes and the
mapping is asymptotic. In other words, the cost of the nervous system adapting to the
middle range is the decreased sensitivity at the ends of the stimuli continuum.

Kahneman and Tversky hypothesized that the same principle applies also to non-physical
attributes such as prestige and wealth: past and present experience define a central range,
to which people habituate and relative to which stimuli (in this case numbers) are
perceived (Kahneman and Tversky, 1979, p: 277). The reference point refers to this
adaptation level while diminishing sensitivity reflects the diminishing impact of a number
as one moves away from the reference point. In the consequence domain for instance, a
difference between a yearly salary of $60,000 and a yearly salary of $ 70,000 has a bigger
impact when current salary is $50,000 than when it is $40,000 (Tversky and Kahneman,
1991). Similarly, the impact of a loss of $10,000 is greater when the reference point is
$40,000 than when it is $50,000.

47
Changes in the same stimulus do not yield necessarily changes of the same nature in the sensation or the
perception. Put differently, the psychological response is a concave function of the magnitude of the
physical change (Kahneman and Tversky, 1979, p: 278)

125
The Localization of the Reference Point
The reference point is usually taken as the current asset position to which outcomes are
evaluated as changes in wealth (Tversky and Kahneman, 1991; Rabin, 2000; and Rabin
and Thaler, 2001).48 It’s influenced nevertheless by many factors, among them, recent
losses, aspirations, or expectations and therefore, may shift from the status quo across
situations. The localization of the reference point is important for the order of preferences
(Kahneman and Tversky, 1979, pp: 286-287)49. For instance, an investor who has
integrated his assets (set the reference point to zero on the scale of wealth) is likely to
choose differently from an investor whose reference point is his current asset position.50
Also an investor who has not adapted to recent losses is likely in the domain of losses to
behave more aggressively.51 The difficulty for financial practitioners to control for losses
is well documented in Glick (1957) and Kleinfield (1983).

Loss Aversion
Loss aversion refers to the asymmetrical treatment of gains and losses relative to a
reference point: outcomes that are perceived as losses are experienced more keenly than
outcomes perceived as gains.

Empirically, Kahneman and Tversky found that most people reject symmetric bets of the
form (-$100, 0.5; $100, 0.5): “the aggravation that one experiences in losing a sum of
money appears to be greater than the pleasure associated with gaining the same
amount…moreover, the aversiveness to symmetric bets increases with the size of the
stakes” (Kahneman and Tversky, 1979, p: 279). Their finding was confirmed later by
numerous empirical studies showing that loss aversion is a major factor in observed risk
aversion (Thaler, 1980, Shefrin and Statman, 1985; Cachon and Camerer, 1996; Gneezy

48
Markowitz (1952) was the first to argue for considering future outcomes as changes of wealth from a
customary level.
49
Schmidt (2003) and Sugden (2003) give general preference axiomatizations for varying reference points
and Bleichrodt, Pinto, and Wakker (2001) use shifts of the reference point to compare preferences at
different reference points, assuming the same utility at the same point for the varying reference points.
50
Let x< y < z < 0,( x , p; z ,1− p )f ( y ,1) shows risk seeking but when the assets are integrated,
( w+ y ,1)f ( w+ x , p;w+ z ,1− p ) shows risk aversion.
51
A person who has just lost 2000 and is facing a choice between a sure gain 1000 and 50/50 chance of
winning 2000 or nothing, is likely to code the choice as between (-2000,0.5) and -1000 rather than a choice
between (2000, 0.5) and 1000 and to prefer the former to the latter (Kahneman and Tversky, 1979, p: 286).

126
and Potters, 1997; Thaler et al., 1997; Bateman et al., 1997; Benartzi and Thaler, 1997;
Payne, Laughhunn and Crum, 1981; Schoemaker and Kunreuther 1982; Hershey and
Schoemaker 1985; Samuelson and Zeckhauser, 1988; Kahneman, Knetsch and Thaler,
1990; Tversky and Kahneman,1991; and Barberis, Huang and Santos, 2001).

Intuitively, to evaluate the attractiveness of possible prospects, the conscious mind

integrates the past and the present subjective imaginings of the future (Shackle, 1955,
Kahneman and Tversky, 1979, Damasio, 2003). The constrained feelings that obtain
seem to be fundamental components of decision making: Damasio et al. (2000) and
Rustichini et al, (in press)52 demonstrate that people who cannot respond emotionally to
the contents of their thoughts show defects in planning and judgment and lack the
aversion to ambiguity or to losses. The feeling that arises in conjunction with an action
resulting in a loss is aggravation accompanied with the desire to avoid a similar situation
enlarging thus the risk which lies at the interface of what one desires and what one wishes
to avoid. 53

The theoretical support for loss aversion came later from Rabin (2000). Observed only in
mixed gambles loss aversion reconciles small risk aversion for small stakes54 which EU
cannot explain, with realistic degrees of risk aversion for higher stakes. Evidence for risk
aversion for small stakes in one shot mixed gambles is given by Samuelson’s (1963)
which shows that individuals reject favorable mixed gambles of the type: ($11, 0.5; $-10,
0.5) and by Kahneman and Tversky’s (1979) which shows that individuals reject
symmetric bets of the type ($11, 0.5; $-11, 0.5). Rabin’s suggestion favoring the use of

52
Antonio Damasio (2000) and his group of neuroscientists undertook a study with people with damage in
the ventro-medial part of the pre-frontal cortex (VMPFC) which showed gross defects in the people’s
planning and judgment despite a high level of performance in language and intelligence tests. (A person
with VMPFC damage cannot respond emotionally to the content of his thoughts). Also, Rustichini,
Dickhaut, Ghirardato, Smith and Pardo (in press) show that people with VMPFC damage show lack of
aversion to ambiguity or losses in situations similar to the Ellsberg Paradox.
53
“The thing I fear most is fear” (Montaigne, 1588, Essais, Book I, 18, ‘De la Peur’). “Fear serves as a
magnet for fear” (Wollheim, 1999, p: 63-65).
54
Officer and Halter’s (1968) shows that even for small amounts of money farmers have non-linear
utilities.

127
loss aversion to explain risk aversion- has been reiterated in “Anomalies: Risk Aversion”
by Rabin and Thaler (2001).

3.1.2 The Utility Function

Because of diminishing sensitivity, the utility function is concave for gains, u " ≤ 0 and
convex for losses, u " ≥ 0 and the reversal of preferences around the reference point is
labeled the reflection effect. The empirical evidence of the phenomenon is listed in
section 2.2.

Kahneman and Tversky (1979) also hypothesize that loss aversion is equivalent to a
utility function that changes abruptly at the reference point. Graphically, the utility
function u is steeper for losses than for gains u '( x) < u '( − x) for x ≥ 0.

Or for all x > y ≥ 0, u ( x ) − u ( y ) ≤ u ( − y ) − u ( − x)

The utility function is thus: (1) defined on deviations from the reference point; (2)
generally concave for gains and convex for losses it has an S shape; (3) steeper for losses
than for gains. The proposed utility function is steepest at the reference point in marked
contrast to the utility function postulated by Markowitz’(1952). Figure 6 shows the
utility function assuming prospect theory.

128
 Figure 6: The Utility Function Assuming Prospect Theory

3.1.3 The Weighting Function

The weighting function has two reference points, certainty and impossibility, which
correspond to the end scales 0 and 1 of the probability domain. Diminishing sensitivity
implies that a difference from 0.55 to 0.6 in probability has less impact than the
difference between 0.05 and 0.1 or between 0.9 and 0.95. The implication is consistent
with the overweighting of extreme probabilities and the underweighting of intermediate
probabilities and explains the Allais paradoxes.55

55
Ranking the outcomes not only eliminated prospect theory’s violations of stochastic dominance but also
fitted well with the diminishing sensitivity principle hypothesized by Kahneman and Tversky (1979).

129
The weighting function is regressive: first it’s greater than p then smaller than p
(intersects the diagonal from above) asymmetric, inverse S-shape (concave first then
convex). Quiggin (1982) has situated the cross over at 0.5 but later studies support a cross
over value between 0.3 and 0.4 (Camerer and Ho, 1994; Tversky and Fox, 1995; Wu and
Gonzalez, 1996, 1998; Prelec, 1998; Abdellaoui, 2000; Bleichrodt and Pinto, 2000). The
weighting function is also subadditive.56

Another important characteristic is its sign dependency: there are two weighting
functions, one for gains and one for losses that are similar in shape but not identical. The
inverted S shape function is modeled as a single parameter function (Tversky and
Kahneman, 1992; Prelec, 1998) or as a two parameter function (Gonzalez and Wu, 1999;
Lattimore et al, 1992) to distinguish the elevation of the function (which refers to the
degree of overweighting/ underweighting) from its curvature (which refers to how people
discriminate intermediate probabilities). Abdellaoui (2000), who estimated the
probability weighting function for the two models, showed that: 1) that when the
weighting function was assumed to have Tversky and Kahneman (1992) single-parameter
form, the median estimates obtained were very close to those obtained by Tversky and
Kahneman (1992) for gains and losses which were 0.61 and 0.69 respectively and 2) that
when the weighting function was assumed to have the Lattimore et al. (1992) two-
parameter function, it exhibited more elevation for losses than for gains. That is the
probability weighting function exhibits less pronounced curvature and more elevation for
losses than for gains. Figure 7 shows the weighting functions for gains and losses
estimated by Tversky and Kahneman (1992).

56
Lower subadditivity is also known as the possibility effect: w( q )− w(0)≥ w( p + q )− w( p ) whenever
p + q ≤1−ε for constant ε ≥0 and upper subadditivity is also known as the certainty effect:
w(1) − w(1− q )≥ w( p + q )− w( p ) whenever p≥ε ' for constant ε '≥0 (Abdellaoui, 2000).

130
 Figure 7: The Weighting Functions for Gains and Losses

3.1 THE FORMAL MODEL

Cumulative Prospect theory applies to objective and subjective and uncertainty. This
section presents the formal model under both.

3.2.1 Decision under Risk

Let X = {x1,…, xn } be a rank-ordered finite set of monetary outcomes including a neutral

outcome 0. To recall, in decision making under risk, a lottery is described by a finite

131
probability distribution over X. It is denoted by (x1,p1; …xn, pn) yielding outcome xj with
probability pj for j = 1…, n; the probabilities are non-negative and sum to one.

Following CPT, the utility of a lottery depends on a utility function u and a probability
weighting function w. The function u defined on gains and losses is a continuous strictly
increasing function satisfying u (0) = 0. The function w is a strictly increasing function
from [0,1] to [0,1] with w(0) = 0 and w(1) = 1.

Furthermore, the weighting function is differentiated for gains and losses into w+ and w-
leading to sign dependence with w+(0) = w-(0) = 0 and w+(1) = w-(1) = 1.

For the lottery L = (x1,p1; …xn, pn) in which x1 ≥… xk ≥ 0 ≥ xk+1 ≥…≥ xn and 0 ≤ k

≤ n, all outcomes are gains if k = n and all outcomes are losses if k = 0. 57

The CPT functionals are represented by equations (3.1), (3.2), (3.3):
n
CPT ( L) = ∑ π j u ( x j ) (3.1)
j =1

Where
u ( xi ) = ( xi )α 58 for positive changes in wealth and

u (xi) = - λ (- xi )β : for negative changes in wealth59 where λ is the loss aversion factor
and α , β the parameters that define the shape of the (Tversky and Kahneman, 1992)
utility function fitted to a power function. The decision weights are defined as follows:

j j −1
π j = w (∑ pi ) − w (∑ pi ) for all j ≤ k
+ +
(3.2)
i =1 i =1

57
The Chateauneuf and Wakker’s (1999) notations for the axiomatization are adopted throughout for
consistency.
58
There is no need in prospect theory to adopt the power function for utility. Tversky and Kahneman fitted
parametrically their findings to the power because preference homogeneity, that is multiplying the
outcomes of prospect by a constant k > 0 multiplies its cash equivalent by the same constant, is both
necessary and sufficient under their theory to represent utility as a power function.
59
A decision maker is well advised to assign a greater weight lamda to negative than to positive outcomes,
to reflect the asymmetry between the experience of gains and losses (Tversky and Kahneman, 1991).
Utility is a ratio scale under CPT, i.e. it’s unique up to a multiplication by a positive constant.

132
 n n
π j = w− (∑ pi ) − w− ( ∑ pi ) for all j > k. (3.3)
i= j i = j +1

π 1 = w+ ( p1 ) for j = 1 and π n = w− ( pn ) for j = n. These decision weights do not

necessarily sum to one.

3.2.2 Decision under Uncertainty

Under uncertainty, the utility of a lottery is represented by (x1, A1 ; x2, A2 ; …xn, An). Aj is
a subset of a state space S called an event, (A1, A2, …An) is a partition of S and xi ∈ X is
the outcome associated with the states contained in Aj

Under CPT, the utility of the lottery (x1, A1 ; x2, A2 ; …xn, An) in which x1 ≥… xk ≥ 0 ≥
xk+1 ≥…≥ xn is given by the following formula:

k n
U (L) = ∑ π i+u ( xi ) +
j =1
∑π
j = k +1
i
−
u ( xi )

Where the decision weights are defined by :

π +j = W + ( A1 ∪ ... ∪ Aj ) − W + ( A1 ∪ ... ∪ Aj −1 )

π −j = W − ( Aj ∪ ... ∪ An ) − W − ( Aj +1 ∪ ... ∪ An )
With π 1 = W + ( A1 ) for j = 1 and π n = W − ( An ) for j = n .

For gains and losses under risk and uncertainty, decision weights as well as subjective
probabilities show subadditivity.60 Research findings show that experts (option traders)
have w(p) = p ∀p ∈ (0,1) because of their familiarity with probabilities and calculus
(Fox, Rogers and Tversky, 1996) while both lay and experts have subadditive probability
judgment (Tversky and Koehler, 1994) i.e. they are not immune to biases and heuristics.

60
Subadditivity under uncertainty is defined as follows: LSAW : ( A)≥W ( A∪ B ) −W ( B ) provided
W ( A∪ B )≤W ( S − E ) ; USA:1−W ( S − A)≥W ( A∪ B )−W ( B ) provided that W ( B )≥W ( E ') where E and E ' are
boundary events (Abdellaoui, Vossmann and Weber, 2005).

133
CPT is axiomatized by Chateauneuf and Wakker (1999) for risk and by Wakker and
Tversky’s (1993) for uncertainty61.

3.3 THE AXIOMATIZATION OF CPT FOR RISK

A central idea to the axiomatization of CPT under uncertainty and under risk is tradeoff
consistency. The latter is also essential to the elicitation method used in this experimental
work to encode the utilities of financial practitioners. Therefore, before presenting it
formally, this section provides first the definition of the idea of tradeoff based on Wakker
(1994). The intuition for tradeoff consistency is then illustrated through an example taken
from Wakker and Tversky’s (1993, p: 149) which axiomatizes in one stroke Savage’s
SEU, Schmeidler’s (1989) and Gilboa’s (1987) CEU (cumulative expected utility), RDU
and CPT. The axiomatization for risk by Chateauneuf and Wakker (1999) follows.

3.3.1 The Idea of Tradeoff

This subsection defines the idea of trade-off as Wakker (1994)62 has adapted it for
decision making under risk. The notation with the star superscript below indicates the
relation is quaternary; it might be interpreted as a revealed ordering of strength of
preferences. The outcomes in the lotteries are rank-ordered.

Definition 3.1:
For four outcomes α , β , γ , δ , we write [α ; β ] ≽* [γ ; δ ] or for short αβ ≽* γδ if
(x1,p1;…; α , p j ; …; xn, pn) ≽ (y1,p1;…; β , p j ; …; xn, pn) and

(x1,p1;…; γ , p j ; …; xn, pn) ≼ (y1,p1;…; δ , p j ; …; xn, pn)

For some j, p1,…yn

Substituting RDU for the preferences at constant probabilistic risk and canceling the
common positive decision weights shows:

61
Other axiomatizations of CPT include: Luce (1991); Luce and Fishburn (1991); Tversky and Kahneman
(1992); Schmidt (2001); Schmidt and Zank,(2001).
62
The author has used it previously in several papers for decision making under uncertainty (e.g. Wakker,
1989a)

134
If RDU holds, then:
α β ≽* γδ ⇒ u (α ) − u ( β ) ≥ u (γ ) − u (δ )

α β f * γδ holds if any of the weak preferences is strict, then if RDU holds, and
α β f * γδ ⇒ u (α ) − u ( β ) > u (γ ) − u (δ )

The ≻* agrees with the ordering of utility differences. In words, the replacement of the
outcomes α , β by γ , δ has caused a reversal of preference. The latter being observed
under constant probabilistic risk must be explained by the change in the outcomes; the
tradeoff α β (receiving α instead of β must be a stronger improvement (less serious loss)
than the tradeoff γδ .

3.3.2 The Intuition for Tradeoff Consistency

The main idea is that if inconsistencies in revealed tradeoff comparisons uncover

deviations from the model assumed, then tradeoff consistency is necessary (it turns out to
be also sufficient for CPT given some natural conditions (Chateauneuf and Wakker,
1999, p: 142)). To illustrate, the following example is taken from Wakker and Tversky
(1993, p: 149).

Consider the lottery (x1, x2) yielding $x1 if state 1 obtains and $x2 if state 2 obtains and
the following pattern of preferences:

(11, 20) ≽ (10, 21) and

(31, 20) ≺ (30, 21)

Applying SEU to both gives:

P1 * u(11) + P2 * u(20) ≥ P1 * u(10) + P2 * u(21) and

P1 * u(31) + P2 * u(20) < P1 * u(30)+ P2 * u(21)

Where P1 * P2 * are the decision maker’s probabilities for states 1 and 2. Under subjective
expected utility (SEU) the following ordering of value differences u(11) – u(10) > u(31)-
u(30) is obtained which implies that receiving 11 instead of 10 is a stronger

135
improvement than receiving 31 instead of 30 and can be interpreted as a revealed

ordering of strengths of preferences: [11; 10] ≻* [31; 30].

If however the following pattern of preferences is also observed:

(40, 31) ≽ (41, 30) and
(40, 11) ≼ (41, 10)

The opposite ordering of utility differences: u(31) − u(30) ≥ u(11) − u(10) is observed
revealing an inconsistency with SEU but not with RDU or CPT where the outcomes 20,
21 are the best outcomes in the first preference and the worst in the second. A decision
maker who pays more attention to the worst outcome will assign more weight to the
outcomes 20, 21 when they are the least desirable outcomes. If the weights attached to
their states are not the same the inference u(11) − u(10) > u(31) − u(30) is not valid
(Wakker and Tversky, 1993, p: 150-151) but the inference u(31) − u(30) > u(11) − u(10)
remains valid because the state 2 yields the least desirable outcomes in both preferences.

Hence, when the two lotteries are comonotonic,63 the decision weights are the same and
cancel. In that case, contradictory inequalities of utility differences are avoided. Under
CPT because sign dependency allows decision weights to differ depending on whether
they are associated with gains or losses, restricting comonotonicity to sign
comonotonicity (the critical outcomes must be either all gains or all losses) is necessary.

3.3.3 The Axiomatization

This section provides first the basic definitions of decision under risk and cumulative
prospect theory. Rank dependence is accommodated by having the outcomes ranked in
the lotteries and sign dependence is accommodated by requiring all outcomes to have the
same sign. It is shown next that the * relations elicit orderings of utility differences under
cumulative prospect theory. Tradeoff consistency is then defined and the theorem stated
formally.

63
Two acts x and y are comonotonic if xi > xj wherever yi > yj for any two states i and j, assuming constant
probabilities throughout.

136
The Basic Concepts
A function U represents ≽ if, for all prospects, P, Q, P ≽ Q iff U (P) ≥ U (Q); then ≽ is
a weak order, complete, transitive. ≽ is continuous if the sets:
{(x1,…,xn ) ∈ X n : (x1, p1;…; xn, pn ) ≽ (y1, p1;…; yn, pn )} and
{(x1,…,xn ) ∈ X n : (x1, p1;…; xn, pn ) ≽ (y1, p1;…; yn, pn )}

are closed sets for every n and fixed n-tuple of probabilities (p1,… ; pn) and lottery (y1
p1;…; yn pn ). Stochastic dominance is satisfied if {(x1 p1;…; xn pn ) ≽ (y1 p1;…; yn pn )}
whenever xj ≽ yj for all j with a strict preference for at least one j with pj > 0.

Tradeoff Consistency

For a lottery P = (x1, p1;…; xn, pn ) and an outcome α , α j P is defined as resulting from
P by replacing xj by α , i.e.

α j P = (x1,p1;…; xj-1,pj-1;…; α , p j ; , xj+1, pj+1;…; xn, pn)

Because the weighting function under CPT is differentiated for gains and losses into w+
and w- leading to sign dependence, it’s crucial that all outcomes have the same sign.
Hence, a prospect is divided into its gain part and into its loss part. For any lottery P
then, the lottery P+ is obtained if all consequences xj < 0 are replaced by 0 and the lottery
P- is obtained if all consequences xj > 0 are replaced by 0.

The following relations can be used theoretically and empirically to elicit the ordering of
utility differences.64 We write [α ; β ] ≽* [γ ; δ ] or αβ ≽* γδ for short if four

outcomes α , β , γ , δ , are all gains or all losses and there exists P = (x1 p1;…; xn pn ) and
Q = (y1 p1;…; yn pn ) with the same probability tuple p1….pn and an index j with pj > 0
such that
α j P ≽ β jQ and
γ j P ≼ δ jQ

64
As shall be seen in Part II, this work uses the * relations to elicit the utilities of financial practitioners.

137
We write αβ f * γδ if in the lower preference we have ≺ instead of ≼
α j P ≽ β jQ and (3.4)
γ j P ≺ δ jQ (3.5)

Substituting the CPT’s formulas described in (3.1), (3.2) and (3.3) in (3.4) and (3.5) and
satisfying and sign-comonotonicity, i.e., as Chateauneuf and Wakker (1999) emphasize,
the decision weights for α j P and γ j P are the same and are written as π i , i = 1,…n, and

the decision weights for β j Q and δ j Q are also the same and are written as λi , i = 1,…n.

The π i ’s may differ from the λi ’s because of different signs of outcomes. The jth

outcomes in all lotteries have the same sign however, hence π j = λ j is given by equation

(3.2) if the outcomes are positive and by equation (3.3) if they are negative. This equality
is crucial and is emphasized by: µ = π j = λ j

∑ π u ( x ) + µu (α ) ≥ ∑ λ u ( y ) + µu ( β ) ; hence
i≠ j
i i
i≠ j
i i

µ (u (α ) − u ( β )) ≥ ∑ λi u ( yi ) − ∑ π i u ( xi )
i≠ j i≠ j

And

∑ π u ( x ) + µ u (γ ) < ∑ λ u ( y ) + µ u (δ ) ; hence
i≠ j
i i
i≠ j
i i

µ (u (α ) − u (δ )) < ∑ λi u ( yi ) − ∑ π i u ( xi ) ; hence
i≠ j i≠ j

µ (u (α ) − u ( β )) > µ (u (γ ) − u (δ ))

α β ≻* γδ ⇒ u (α ) − u ( β ) > u (γ ) − u (δ ) (3.6)

αβ ≽* γδ ⇒ u (α ) − u ( β ) ≥ u (γ ) − u (δ ) (3.7)

Tradeoff consistency holds if there are no outcomes α , β , γ , δ

such that both αβ f *γδ and γδ ≽* αβ

138
Formally, the above results can be stated in the following theorem.
Theorem: The following statements are equivalent:
(i) Cumulative Prospect theory holds with a continuous value function
(ii) ≽ Satisfies the following conditions:
(a) Weak ordering
(b) Continuity
(c) Stochastic dominance
(d) Tradeoff consistency65

To recapitulate, cumulative prospect theory, a descriptive decision making model for risk
and uncertainty, hypothesizes a neutral reference point relative to which future outcomes
are evaluated as gains or losses rather than increases or decreases in total wealth. This
notion, the cornerstone of the theory came to the authors from observations that people
are in general risk averse for gains and risk seeking for losses and that they are extremely
reluctant to accept mixed prospects.66 The perception of a reference point implies 1)
diminishing sensitivity for the two part cumulative functional of the valuation rule and 2)
loss aversion. These two components of psychological nature are invoked by Tversky and
Kahneman (1992) to explain the characteristic reflection pattern of attitudes towards risk
in terms of the utility function and the probability weighting functions. The following
section is devoted to the characterization of risk under EU, RDU, and CPT.

3.4 THE CHARACTERIZATION OF RISK ATTITUDE

An important property of expected utility theory is that different notions of risk aversion
defined independently of any model are equivalent to the concavity of the vNM utility
function that is to the diminishing marginal utility principle.

65
Sign-dependence is accommodated by the requirement that all outcomes should have the same sign.
66
The behavior has been documented in a review by Fishburn and Kochenburger (1979) around the time
they wrote their first version of the theory. Fishburn and Kochenberger’s is a review of five independent
studies by Barnes and Reinmuth (1976), Grayson (1960), Green (1963), Halter and Dean (1971), and
Swalm (1966).

139
Yaari (1987) has however shown that “risk aversion and diminishing marginal utility of
wealth are horses with different colors” and an agent can be risk averse without a
concave utility. Moreover, Chateauneuf and Cohen (1994) have shown that an agent can
be risk averse even with a convex utility provided his probability weighting function
synonymous to probabilistic risk aversion is sufficiently convex.

These two components of risk aversion could be further differentiated as suggested by the
two “empirically desirable”67 generalizations of rank dependent utility by Tversky and
Kahneman (1992) who brought in loss aversion and allowed for different risk attitudes
for gains than for losses to explain the complex patterns of risk behavior displayed in
even very simple contexts.

The following section presents the behavioral notions of risk aversion and establishes the
links between these notions and the concavity/convexity of the utility function. The
characterizations of risk aversion under EU, RDU and CPT are presented next.

3.4.1 Notions of Risk Aversion

There are many ways to define risk aversion. Most notably, Rothschild and Stiglitz’
(1970) definition of strong risk aversion based on the notion of an increase in riskiness
and Pratt’s (1964) and Arrow’s (1965), notion of weak risk aversion which represents
more the intensity of risk aversion than risk aversion in the strict sense.

Strong Risk Aversion

The notion of an increase in riskiness is directly linked to the notion of second order
stochastic dominance (SSD). The formal definition of first order stochastic dominance
F ≥1 G (the cumulative distribution of F is uniformly below the cumulative distribution of

67
Köbberling and Wakker (2005)

140
G) was given previously and it can be shown that if F ≽ G according to FSD then
necessarily that implies that the expected utility of F is greater than the expected utility of
G for any agent with a monotonic non-decreasing utility function.68 All of expected
utility, rank dependent utility, dual theory and cumulative prospect theory satisfy first
order stochastic dominance.

SSD ranks cumulative distributions in terms of their relative riskiness that is in terms of
the spread of their probability mass when neither distribution is uniformly below the

other. If F ≽ G according to SSD (the probability mass under F is less spread out than it
is under G), then necessarily that implies that the expected utility of F is greater than the
expected utility of G for any agent with a non-decreasing concave utility function.69 Thus
FSD implies SSD but not vice versa. Formally, second order stochastic dominance is
defined as follows:

Definition 3.2:
∀ F, G ∈ D(X), a lottery F is defined to dominate a lottery G by second order stochastic
dominance (F ≥2 G) if G(x) –F(x) ≥ 0 ∀x ∈ X

If, in addition, F and G have the same mean (E(F) = E(G)) it’s said that G is a mean
preserving increase in spread (MPIS) constructed by moving probability mass away from
the center of the distribution to its tail in such a manner that the mean remains the same.
The following definition of strong risk aversion by Rothschild and Stiglitz (1970)
delineates a risk averse individual as someone who always dislikes mean–preserving
spreads.

Definition 3.3:
An agent has strong aversion for risk if for any two lotteries F and G having the same
mean such that F dominates G by second order stochastic dominance, he prefers F to G.

∀ F, G ∈ D(X), E (F) = E (G), F ≥2 G ⇒ F ≽ G.

68
For proof of the FSD rule cf. (Hadar and Russell, 1969; Hanoch and Levy, 1969; Rothschild and Stiglitz,
1970)
69
Ibid for proof of the SSD rule.

141
Weak Risk Aversion

Risk aversion can be also defined in terms of the certainty equivalent that is the amount
of cash one is willing to accept with certainty in lieu of facing a lottery P. An agent is
indifferent between the lottery P and its certainty equivalent lottery which is the sure-
thing lottery that yields the same utility as the lottery P, i.e. δ CE ( P ) ∼ P; however, the

certainty equivalent CE (P) which is the inverse of the utility of the lottery varies
according to the shape of the utility function as can be seen from Figure 8 in Appendix A.
When the shape of the utility function is concave, the certainty equivalent is less than the
expected value E (P). The difference denoted by π ( P ) = E ( P ) − CE ( P ) is known as the
risk premium i.e. the maximum amount one is willing to forego in order to obtain an
allocation without risk (Pratt, 1964). Ad modum, when the shape of the utility function is
convex, CE (P) > E(P) and when it is linear, CE (P) = E(P) (Eeckhoudt and Gollier,
1992, p:26-28).

An agent is risk averse if CE (P) < E(P) or π ( P ) > 0

An agent is risk seeking if CE (P) > E(P) or π ( P ) < 0

An agent is risk neutral if CE (P) = E(P) or π ( P ) = 0

Definition 3.4:
An agent has weak risk aversion if for any lottery P ∈ P he prefers to this lottery the

certainty of its expectation: ∀ P ∈ P, E (P) ≽ P

It’s possible to order the notions of risk aversion. Strong risk aversion implies weak risk
aversion but not vice-versa. The following section characterizes risk aversion under the
different theories.

142
3.4.2 Characterization of Risk Aversion under the Different Theories

The following section establishes the links between the notions of risk aversion defined
independently of any model presented above and the convexity and concavity of the
utility function under the different theories. That is it characterizes risk aversion under
EU, RDU and CPT respectively.

The Characterization of Risk Aversion under EU

Proposition 3.1:
If an agent satisfies the hypotheses of the expected utility model, then the three
statements are equivalent:
i) The agent has a strong aversion for risk
ii) The agent has a weak aversion for risk
iii) The agent’s utility function is concave

Arrow (1964) and Pratt (1965) were able independently to find a measure for the degree
of risk aversion in terms of the properties of the utility function by approximating the risk
premium of a lottery. Using approximation formulas of a continuous and differentiable
1 u "( w)
function, they showed that under EU, the risk premium is equal to − σ 2 where
2 u '( w)
−u "( w)
σ 2 is the variance of the lottery and is the measure of risk aversion.
u '( w)
−u "( w)
This local measure ( ) known as the degree of absolute risk aversion is
u '( w)
fundamentally specific of the individual at a certain level of wealth and enables the
comparison of degrees of risk aversion between two individuals who might have the
same wealth, might be endowed with the same lottery and yet demand different risk
premiums because the curvatures of their utility functions are different. This measure of
the intensity of risk aversion is independent of the notions of risk aversion defined above
and does not carry over to the other models. The approximation gives different results
under different models (Eeckhoudt and Gollier 1992, p: 26-27).

143
The Characterization of Risk Aversion under RDU

The characterization of risk under EU excludes thus the possibility of risk seeking
behavior for a decision maker exhibiting a concave utility function. However under rank
dependent utility theories, the decision maker’s behavior is characterized by two
functions u and w which could explain a mixture of risk seeking and of risk aversion
(Quiggin, 1991). Yaari (1987) having shown that an agent with a convex probability
weighting function can be risk seeking without having a convex u, Chateauneuf and
Cohen (1994) further demonstrate that an agent can be risk averse even if he has a convex
u provided a weak definition of risk aversion is adopted. The provision is important for
when a strong definition of risk aversion is considered, global risk seeking behavior is
inconsistent with the diminishing marginal utility of wealth as shown by Chew, Karni and
Safra (1987). The latter defined strong risk aversion for RDU as follows:

Proposition 3.2:
An agent satisfying the hypotheses of rank dependent utility has a strong aversion for risk
iff his probability weighting function w (.) is convex and his utility function u (.) is
concave; w(.) and u (.) are differentiable

A convex weighting function can be interpreted as a form of pessimism.70 The intuition

is developed in Diecidue and Wakker (2001)71. The decision maker may decide
deliberately and consciously that the best/worst outcomes should receive more attention,
more importance weight than the intermediate outcomes; hence under pessimism
improving the ranking of the outcome decreases the decision weight. A weak
probabilistic risk aversion has been defined by Quiggin (1982) and Yaari (1987) as
follows:

70
Assume a lottery yields outcome x with probability p. Let q denote the ranking position of x, i.e. the
probability of receiving a lower or equal outcome. The decision weight of x then is w (p+ (1-q)) – w(1-q)
which is under pessimism decreasing in q iff w is convex (Diecidue and Wakker, 2001, p: 288).
71
Psychology’s contributions to the intuition of rank dependence comprise Birnbaum (1974), Lopez (1987)
independently of Quiggin, Weber (1994) and others and are listed in Diecidue and Wakker (2001).

144
Proposition 3.3:
An agent satisfying the hypotheses of RDU has weak risk aversion iff for all p ∈ [0, 1],
w (p) ≤ p; u (.) is differentiable and concave

Probabilistic aversion in the weak sense gives a necessary and sufficient condition of risk
aversion under the assumption of diminishing marginal utility. Chateauneuf and Cohen
(1994) provide a sufficient condition for weak risk aversion under the assumption of
increasing marginal utility as follows:

Proposition 3.4:
An agent satisfying the hypotheses of rank dependent utility and having a convex u (.)
can be risk averse in the weak sense if his function w(.) is sufficiently convex, i.e if he is
sufficiently pessimistic

For strong risk aversion, u (.) is concave, and w (.) convex as in (Chew, Karni and Safra,
1987), that is a pessimistic agent with a concave u (.) is universally risk averse.

The links established thus far between the different notions of risk aversion and the
preference functionals under RDU can be extended to dual theory a special case of RDU.
For CPT, the results obtained under RDU can be extended to prospects restricted to either
gains or losses but not to mixed prospects.

The Characterization of Risk Aversion under CPT

CPT a reference dependent model has three distinct notions of risk: an intrinsic utility, a
probability weighting function and loss aversion. Since loss aversion is only observed in
mixed prospects the characterization of risk aversion under CPT distinguishes among
positive prospects (all gains), negative prospects (all losses) and mixed prospects.

The conditions for strong risk aversion under CPT and RDU (the special case of CPT
where the weighting function for losses is the dual of the weighting functions for gains,
i.e., w− ( p ) = w+ (1 − p )∀p ∈ (0,1) coincide if the prospects considered are only positive or
only negative.

145
 Proposition 3.5:
For either gains or losses an agent satisfying the hypotheses of CPT has a strong aversion
for risk iff his probability weighting function is convex and his utility function is
concave

For mixed prospects, risk aversion under CPT is characterized through a joint condition
on utility curvature, probability weighting and loss aversion by Schmidt and Zank (2002).
Their characterization supports the Köbberling and Wakker (2005) index of loss aversion
which is the first axiomatically founded index proposed. Table 6 summarizes the
different characterizations of risk aversion under the different theories.

Table 6: Characterization of Risk Aversion under the Different Theories

EU DT RDU CPT CPT CPT

(+) (-)
Prospects Prospects M. Prospects
u(.) w(.) u(.) , w(.) u(.) , w(.) u(.) , w(.) u(.), w(.), LA
Strong u (.) w (p) u (.) u(.) u (.)
Aversion concave convex concave, concave, concave, Sufficient
w(p) w (p) w(p)
convex convex convex Condition
u (.)
Weak concave w(p) ≤ p Sufficient Sufficient Sufficient Sufficient
Aversion Condition Condition Condition Condition

146
COMMENTS AD FINEM ON PART I

Rank dependent utility theories explain systematic patterns termed paradoxical under EU
and refine the understanding of risk aversion by accounting for observed risk behavior
that is inconsistent with the characterization of risk under EU. The refinement is twofold:
1) under RDU where all notions of risk aversion are not confounded with a concave
utility function, an individual with an increasing marginal utility can be risk averse in the
weak sense of Chateauneuf and Cohen (1994); 2) RDU theories distinguish between two
behaviors of different nature which are not discernable under EU. The different rationales
are clearly differentiated in RDU’s treatment of the portfolio selection problem where it
is the custom theoretically and in practice to assume the presence of a risk free security
and to conduct the selection and analysis of the portfolio in two phases: the first is the
scale or the allocation between a riskless and risky assets and the second is the mix or the
allocation within the category of risky assets.72

Yaari (1987) shows that between a riskless asset and a portfolio of risky assets, a risk
averse investor (a pessimistic investor) under DT does not diversify but stays put until
plunging is justified (Yaari, 1987, p: 10), as opposed to a risk averse investor under EU
who will always diversify (Markowitz, 1952; Tobin, 1958). Yaari (1987) also conjectures
that within the portfolio of risky assets, a risk averse investor diversifies according to the
maximin criterion where minimizing the worst result is independent of the utility.

Gayant (2004) proposes deux logiques de décision différentes for the scale and the mix:
between a risky asset and a riskless asset, risk aversion is due to a desire for partial
security reflecting the investor’s unwillingness to eliminate an opportunity for gain while
within a portfolio of risky assets risk aversion is due to a desire for full security through
minimizing the worst outcome. His findings that the maximin criterion is “standard”
under DT and “at the limit” under EU the two limit cases of RDU which isolate the

72
The division is called the separation theorem: “breaking down the portfolio selection problem first among
and then within asset categories seems to be a permissible and perhaps even indispensable simplification
both for the theorist and for the investor himself” Tobin (1958). The division finds its parallel in practice.
All fund managers interviewed distinguish between selection among asset categories and selection within a
category.

147
effects of utility and probability transformation, support Yaari (1987) and confirm that
“at the minimum utility cannot exclusively represent risk aversion”.

The understanding of risk attitude is further refined under CPT: risk attitude has three
components that are affected jointly by a gamble: a reference dependent utility function,
loss aversion and a probability weighting function. The reason is that in addition to
retaining the RDU’s mathematical precision precious to decision theorists, CPT provides
a convenient way of modeling the influence of pervasive phenomena like loss aversion
and the reflection effect on choice.

Research in behavioral finance (Thaler, 1993; Kahneman and Tversky, 2000) indicates
that in financial markets, the control of losses is the major problem for investors (or their
planners).73 The difficulty for controlling losses stems in a large part from the evaluation
of possible prospects as gains or losses relative to a reference point rather than positing
the decision problem in terms of final wealth as in the normative EU. Preferences
indexed to a reference point, as in CPT explain the tendency for risk seeking in the
domain of losses and the difficulty because of loss aversion to perceive the sale price of a
stock independently of its purchase price ending in lower earnings. The many studies
included in Thaler (1993) also point to the importance on decision making of
occupational factors that is psychological factors that seem to unfold in the practitioner’s
environment. Moreover, Barberis, Huang and Santos (2001) find that both loss aversion
and variation of loss aversion with the past movement of the stock market are necessary
to explain the equity premium puzzle.

Thus, in addition to the theoretical advantages that characterize rank dependent theories,
CPT’s empirical realism argues in favor of its use as a good approximation of the
underlying preference function the investor is assumed to optimize leaving it up to the
measurement of the theory’s parameters in the field to show whether the approximation is
good.

73
During the interviews where the financial practitioners were also asked about their major problems the
control of losses emerged as the number one preoccupation for all.

148
This work proposes to use the method Abdellaoui et al. (2006) made available to elicit
under prospect theory and non-parametrically i.e. without any assumptions of the form of
the functional, the utility functions of practitioners in the field of finance. It also proposes
in Experiment II to use a different method to infer the preferences of MBA students.

Part II thus is composed of two experiments. Experiment I investigates the fundamental

preferences of financial practitioners by eliciting the utility function completely and
measuring the degree of loss aversion at the individual level using Abdealloui et al.’s
method. Experiment II investigates qualitatively the preferences of MBA students,
potential financial practitioners without eliciting the utility function using Baucells and
Heukamp (2006). Part II thus consists of a general introduction and two chapters. Each
chapter is devoted to one experiment where the method and the experiment per se are
described, and the results discussed in light of the recent literature. Part II also addresses
the problems encountered such as the need for an agreement on loss aversion’s definition,
the difficulty of empirically disentangling all three components of risk attitude and the
possible influence occupational factors could have on the practitioners’ risk attitude in the
dynamic context of the stock market. A general conclusion follows leading to some final
remarks and direction for future research.

149
PART II: EXPERIMENTAL INVESTIGATION OF INDIVIDUAL
UTILITY FUNCTIONS AND LOSS AVERSION

Part I has shown much evidence that people willingly violate expected utility theory and
that cumulative prospect theory explains most of the violations. Indeed, its authors built it
to fit the individual-level data they have gathered from their experiments motivated as
they were by the observation of a “remarkable discrepancy” in the literature on
preferences between gambles: “the theoretical analysis implied that the carriers of utility
were states of wealth but the outcomes were always described as gains or losses”
(Kahneman, 2000, p: ix). The observation eventually led them to a utility function that is
concave for gains and convex for losses (S-shape) and steeper for losses than for gains.
Part II’s concern is whether the psychologically plausible S-shape is approximately true
and useful.

Many empirical studies with the same concern have confirmed the latter shape for the
utility function. However, most74 of the elicitation methods used across these studies have
assumed specific parametric forms for the utility function making thus the inferences
about these functions dependent on the choice of the functionals.

Although Abdellaoui (2000)75 was the first paper to elicit cumulative prospect theory
model non-parametrically under risk by means of the tradeoff method76, it’s Abdellaoui
et al (2006) that have made available a non-parametric method to elicit the utility
function for gains and losses simultaneously allowing the measurement of loss aversion a
gain/loss exchange rate77. The availability of the method and the ease of its applicability

74
Wu and Gonzalez (1996, 1998) elicit the utility function without any parametric assumption by testing
preference condition but the approach is demanding.
75
Abdellaoui et al. (2005) elicits CPT under uncertainty without any prior knowledge of probability or
utility. Both do not include the measurement of loss aversion.
76
Initially proposed by Wakker and Deneffe (1996).
77
So far, loss aversion has not been separated from the utility function either empirically or theoretically.
Köbberling and Wakker (2005) disentangle loss aversion from the utility function under a “severe
restriction”

150
in a relatively short time offered the possibility to “spend some time in the wild”
specifically in the stock market where time is an expensive commodity.

The method was used hence to elicit non-parametrically and simultaneously the utility
functions of financial practitioners for gains and losses and to measure their individual
loss aversion allowing hence a test of prospect theory in the field. The results were to be
compared subsequently to those obtained in previous studies which have however
investigated mainly students’ preferences. Another experiment was also undertaken to
infer the preferences of MBA students, potential financial practitioners using this time a
newly developed method based on stochastic dominance criteria which Baucells and
Heukamp (2006) have made available.

Probability weighting is not a problem in either method albeit for different reasons, the
two being based on different concepts. The robustness against probability weighting is
fundamental if the method is to remain valid for RDU and CPT and to be applied in
prescriptive decision analysis. Indeed it has been shown that probability weighting is a
major cause of violations of EU and of systematic inconsistencies among different
methods of utility elicitation that should yield the same result under EU (Fischoff (1982);
Hershey and Schoemaker (1985); McCord and de Neufville (1986); Wakker and Deneffe
(1996); Schkade (1988); Bleichrodt et al. (2001)).

Part II, in sum, consists of two experiments aiming at examining individuals’

fundamental preferences under risk in the specific domain of finance: Experiment I uses
Abdellaoui et al.’s (2006) method in the field and Experiment II uses Baucells and
Heukamp’s (2006) in the lab. To each, a chapter is devoted which begins with a brief
introduction showing the gradual yet portentous development of the method from its
predecessors.

In chapter IV, the utilities of practitioners are elicited non-parametrically and

simultaneously for gains and losses and their loss aversion coefficients are measured.
Section 1 presents the elicitation procedure. Section 2 describes the design of the

151
experiment. Section 3 is concerned with the data analysis which includes the non-
parametric elicitations, the parametric fitting and the measurement of loss aversion in the
aggregate as well as the individual level. Section 3 and 4 present the results related to the
practitioners’ shape of the utility function and their loss aversion respectively. The results
in the field are then contrasted with the results of Abdellaoui et al. (2006) in the
laboratory and those of other previous studies. A final section concludes.

In chapter V, the utility functions of MBA students are elicited from lotteries based on
stochastic dominance criteria developed by Baucells and Heukamp (2006). Section 1
presents the intuition for the stochastic dominance conditions for clarification followed
by the characterization of preferences; section 2, the experimental application that is the
objective, and the source of data, section 3 consists of the analysis of data and section 4
and 5 give the results pertaining to the shape of the utility function and loss aversion of
MBA students respectively. A final section concludes followed by a general conclusion
and some remarks and directions for future research.

152
CHAPTER IV. EXPERIMENT I.

FIELD INVESTIGATION USING A PARAMETER-FREE METHOD

FOR THE ELICITATIONS

In so far as classification is needed before evaluation, Fishburn (1967) provides a list and
a classification of 24 methods for estimating utility and Peter Farquhar (1984) describes
the state of the art in utility assessment 17 years later reviewing hence many methods that
were not examined in Fishburn (1967). Farquhar classifies the methods under 1)
preference comparison methods, 2) probability equivalence methods, 3) value
equivalence methods and 4) certainty equivalence methods. The methods however turned
out in time not to be equivalent under EU, the inconsistencies among them shown to be
systematic rather than random undermining the validity of the method’s use in
prescriptive decision analysis.

Hershey and Schoemaker (1984) among others provide evidence of “serious

discrepancies” between the two most common methods used for elicitation the certainty
equivalent method and the probability equivalent method. To correct for the use of the
former as well as for the chaining of responses and the lack of control over ranges and
reference points, McCord and Neufville’ (1984) propose the lottery equivalent elicitation
method. Their procedure although simple, suffers under EU from probability weighting
problems.

The misperception of the latter being an important violation of EU, Wakker and Deneffe
(1996) develop the tradeoff method from the saw tooth method, one of 24 methods listed
and classified by Fishburn (1967). The method uses the same probability across the
various lotteries completely eliminating the distortions of utility measurement which are
due to probability weighting. Abdellaoui’s (2000) and Abdellaoui et al.’s (2005) show
that it can elicit probabilities indirectly assuming CPT under risk and under uncertainty

153
respectively and Bleichrodt et al. (2001)78 suggest its use to correct the certainty
equivalent and the probability equivalent for probabilistic weighting. The method is also
“well-suited” for the axiomatization of CPT under risk (Chateauneuf and Wakker, 1999)
and under uncertainty (Wakker and Tversky, 1993) as was shown in section 3.3 for the
former.

Abdellaoui et al. (2006) develop a new elicitation method for prospect theory. Their
elicitation procedure encodes the utilities for gains and losses simultaneously allowing
hence, the measurement of loss aversion non-parametrically. Although the method itself
could as well be used joint with matching, the way it was implemented in their
experiment was choice-based. Previous studies have found that inferring indifferences
from a series of choices leads to fewer inconsistencies than asking subjects directly for
their indifference values,79 and according to Tversky, Sattath and Slovic (1988), when
subjects are asked to reconsider inconsistent answers they modified the matching in the
direction of the choice.80

A drawback of the tradeoff method is that the utility function elicited is not independent
of a reference point relative to which outcomes could be perceived as gains or losses.
Another, is that it allows error propagation if the assessment of the first utility is not well
done, since the measurements are chained and later responses are based on previous ones.
Nevertheless, both Bleichrodt and Pinto (2000) and Abdellaoui, Vossman and Weber
(2005) who have used similar elicitation methods and have examined in detail the effect
of error propagation on chained measurements found it to be negligible.

78
Bleichrodt et al. (2001) suggest the tradeoff to correct quantitatively the certainty equivalent for
probabilistic transformation and the probability equivalent for loss aversion and probabilistic
transformation rather than qualitatively as in Hodgkinson et al (1999) or in Payne et al. (1999) or in Arkes
(1991).
79
Luce (2000) provides a review of these studies.
80
This observation suggests that choice and matching are both biased in opposite directions (the primary
dimension may be overweighted in the former and underweighted in the latter, the answers reflecting
perhaps a routine compromise rather than the result of a critical reassessment (Tversky, Sattath and Slovic,
1988)

154
Chapter IV consists of five sections. Section 1 describes the background for the
elicitation procedure used based on Abdellaoui et al. (2006). Section 2 describes the
procedure itself and its experimental application. Data analysis follows in section 3. The
results related to the shape of the utility function and to loss aversion are presented next
in sections 4 and 5 respectively. A final section concludes.

4.1 THE BACKGROUND

The Abdellaoui et al.’s (2006) elicitation procedure is a parameter-free method to
completely elicit the utility function under prospect theory. It allows thus the
measurement of loss aversion at the individual level without making any parametric
assumptions. This section reviews in brief the main features of prospect theory to provide
a background for the following section, which describes the elicitation procedure and
applies it in the field.

In the experiment, the practitioner is asked to choose between two lotteries with at most
two distinct outcomes. The discussion is thus restricted to such lotteries; nevertheless, the
estimations are valid for both Kahneman and Tversky (1979) and Tversky and Kahneman
(1992) which coincide on the two-outcome domain. Outcomes are monetary and are
expressed as changes from the reference point, that is, as gains or losses. The reference
point is assumed to be zero and all outcomes mixed and non-mixed are assumed to be
rank-ordered. The individual evaluates each lottery and chooses the lottery that offers the
highest overall utility. The overall utility is expressed in terms of three functions: a
probability weighting function w+ for gains, a probability weighting function w− for
losses and a utility function U. The functions w+ , w− assign a probability weight to each
probability. They are strictly increasing and satisfy w+ (0) = w− (0) = 0 and

w+ (1) = w− (1) = 1 . The utility function U assigns a real number to each outcome, which
reflects the desirability of that outcome. The function U is increasing and satisfies U(0) =
0. U is a ratio scale, i.e. the unit of the function is arbitrarily chosen.

155
The evaluation of the lottery (x, p; y) depends among other things on the sign of the
outcomes x and y. For a non-mixed lottery that is involving only gains or only losses, the
utility computed by:
wi ( p )U ( x) + (1 − wi ( p ))U ( y ), (4.1)
Where i = + for gains and i = - for losses.

For a mixed lottery it is computed by: (4.2)

w+ ( p)U ( x) + w− (1 − p ))U ( y ) .

Kahneman and Tversky (1979) assumed that the probability weighting functions for
gains and losses overweight small probabilities and underweight moderate and high
probabilities, giving rise to an inverse S-shape probability weighting function. The utility
function is assumed to be concave for gains and convex for losses and steeper for losses
than for gains.

4.2. THE EXPERIMENT

The procedure used in the experiment is based on the elicitation of utility midpoints
pointed at before in Köbberling and Wakker (2005) and which have been often used in
axiomatizations of decision models, and it is noteworthy that it requires few
measurements to elicit a given number of utility midpoints.

4.2.1. THE ELICITATION PROCEDURE

The procedure consists of four steps and is summarized in Table 7 below. The second
column describes the quantity assessed, the third the indifference that is sought and the
fourth the implication of this indifference under prospect theory. The final column shows
the choices for the variables that have to be specified. The font size in Table 7 is smaller
than in the other tables in order not to overcrowd the page.

156
 Table 7: Four-Steps Elicitation Procedure
Assessed
Quantity Indifference under Prospect Theory Choice Variables
Step 1 L1 (L1, p’; L*) ∼ (L0, p’; L) p’ = 0.33
L2 (L2, p’; L*) ∼ (L1, p’; L) U(L0) - U(L1) = U(L1) -U(L2) L* =-100
pl L1 ∼ (L2, pl ; L0) w-(pl) = 0.5 L=-600, L0 = -1000
G1 (G1, p’; G) ∼ (G0, p’; G*) p’ = 0.33
G2 (G0, p’; G) ∼ (G1, p’; G*) U(G2) - U(G1) = U(G1) - U(G0) G* = 600
+
pg G1 ∼ (G2, pg; G0) w (pg) = 0.5 G = 100, G0 = 1000
Step 2 Lr ∈ [ L1 , 0] Lr ∼ (LA, pl ; LB) U(Lr) = 0.5 U(LA) + 0.5 U(LB) L1 = -100000
-
Step 3 l Ls ∼ (l, 0.5 ; 0) w (0.5)U(l) = -s s = 0.25
g 0 ∼ (g, 0.5 ; l) w+(0.5)U(g) = s
Gs Gs∼ (g, 0.5 ; 0) U(Gs) = w+(0.5)U(g) = s
Step 4 Gr ∈ [ 0, Gs] Gr ∼ (GA, pg ; GB) U(Gr) =0.5 U(GA) +0.5 U(GB)

The first step, a central step in the procedure for decision making under risk, is the
elicitation of probabilities that have a decision weight of 0.5.81 The elicitation of the
probability for gains or for losses for which w+ ( pg ) = 1/ 2 and w− ( pl ) = 1/ 2 requires

three measurements, that is, three indifferences. Once this is known however, just one
measurement is needed for the determination of a utility midpoint.

For the elicitation of pl a sequence of losses L0, L1 and L2 that are equally spaced in terms
of utility, i.e. U(L0) – U (L1) = U(L1) – U(L2). More specifically, a probability p’ is
chosen and three losses L*, L, and L0, with L* > L > L0. Then losses L1, L2 are elicited
such that a subject is indifferent between the lotteries (L1, p’; L* ) and (L0, p’; L ) and
between the lotteries (L2, p’; L* ) and (L1, p’; L ). Because L* > L, L2 < L1 < L0 is a must.

Under prospect theory, the indifferences (Li+1, p’; L*) ∼ (Li, p’; L), i = 0, 1 imply that

81
For decision under uncertainty, the elicitation of events that have a decision weight of 0.5 can be
interpreted as a generalization of Ramsey’s (1931) “ethically neutral events” (events with subjective
probability of 0.5) under expected utility to prospect theory (Abdellaoui et al., 2006).

157
 1 − w− ( p )
U (L ) − U (L )= (U ( L*) − U ( L), i = 0,1 . (4.3)
i i +1 w− ( p )

Because the expression on the right hand side is constant, it follows that: U(L0) – U (L1) =
U(L1) – U(L2). Hence, L1 is a utility midpoint of L0 and L2. This procedure for eliciting
utility have been pointed out previously by Abdellaoui (2000) and Köbberling and
Wakker (2003).

Having elicited L1 and L2 the probability pl is determined that makes the subject
indifferent between L1 for sure and the prospect (L2, pl; L0). Under prospect theory,
U ( L1 ) = w − ( pl )U ( L2 ) + (1 − w− ( pl ))U ( L0 ). L1 being the utility midpoint of L0 and L2 gives

w− ( pl ) = 1/ 2 . The elicitation of pg is similar except that now three monetary gains G0 >
G* > G are fixed beforehand.

In the second step of the elicitation determines the utility for losses is determined by
eliciting utility midpoints. Once pl is known the utility midpoint of any two losses LA and
LB can be measured by eliciting one indifference only as is shown in step 2 in Table 7.
U( L1) is set equal to −1 for some L1 < 0 which is allowed by the uniqueness properties
of the utility function in prospect theory. The outcome L0.5 is then determined such that
the subject is indifferent between L0.5 for sure and the lottery (L1, pl; 0). Under prospect
theory, this indifference implies that U (L0.5) = −0.5 . The utility is then elicited on the
interval [L1, 0]. For example, by setting LA = L0.5 and LB = 0, the outcome L0.25 is elicited
for which U(L0.25) = −0.25.

The third step is the crucial step in the measurement of loss aversion. In it, the utility of
losses is linked to the utility of gains by eliciting three indifferences. In the first
indifference one of the outcomes elicited in step 2 Ls is chosen and the loss l is
determined such that the subject is indifferent between Ls and (l1, 1/2; 0). It follows that
w− (1/ 2)U (l ) = − s. The second indifference determines the gain g that makes the subject

indifferent between 0 for sure and the lottery (g, 1/2; l). It follows that w+ (1/ 2)U ( g ) = s.

158
The gain Gs that the subject finds equivalent to the lottery (g, 1/2; 0) has then the utility s
and is the mirror image of Ls in terms of utility.

The fourth and final step of the elicitation procedure determines the utility for gains. As
for losses, the probability pg allows to measure the utility midpoint of any two gains GA
GB by eliciting just one indifference. The utility midpoint of and Gs and 0 is determined
first and then the utility for gains on the interval [0, Gs].

4.2.2 THE EXPERIMENTAL APPLICATION

The procedure outlined above was used for the complete elicitation under prospect theory
of the utility of financial practitioners. It was possible thus, to test in the field the theory’s
assumption that the shape of the utility is concave for gains and convex for losses and to
measure the practitioners’ individual degrees of loss aversion, according to the various
definitions that have been put forward in the literature.

The General Setup

Forty six practitioners participated in the study. Most were financial advisors responsible
for managing the portfolios of their respective clients, some however were money
managers, that is, portfolio managers in whose funds, the financial advisors invest a
fraction of their clients’ wealth. They were all affiliated with multi-national financial
institutions. The interviews were obtained through personal contact and were conducted
individually on a computer in their respective offices: Cleveland-Ohio, Boston-
Massassuchets, Manhattan-NewYork, Atlanta-Georgia, Phoenix-Arizona and Beirut-
Lebanon. There were no systematic differences among them and their data was pooled.
Their median age is 40. During the interviews, the practitioners were encouraged to go on
their own pace and were told that there were, no right or wrong answers. The session
took an average of 30 mn. Table 8 in the Appendix has the names of the practitioners’
institutions and their locations.

159
Except for the source of the data, and the type of interaction between interviewer and
interviewee, the general setup and the details of the experiment were the same as in
Abdellaoui et al. (2006). The outcomes were in dollars and they were substantial in order
to be able to detect the curvature of the utility function, utility being approximately linear
over small intervals (Wakker and Deneffee, 1996). The practitioners were not directly
asked for the specific outcome value leading to indifference, previous studies having
shown that inferring indifferences from a series of choices leads to fewer inconsistencies
than asking subjects directly for their indifference value. Instead, every value was
assessed through a series of binary choice questions.

82
The question was framed as follows: The following hypothetical choices are designed
to investigate your attitude towards risk. Please take your time and try to predict your
choices as accurately as possible. The responses are anonymous and there is no correct
answer, hence, no reason not to state your true preference. However, it was pointed out to
the practitioner that the money used for the investment was not his own nor the client’s
which could vary widely from one to another83 but the company’s. In this way, the
choices reflected the preferences of the practitioner as a professional.

The following display illustrates the first choice as it was presented to the practitioner on
the computer screen. The two lotteries A and B corresponding to (L1, p’; L*) and (L0, p’;
L) where L0, p’ and L* are fixed, are displayed as pie charts with the sizes of the pieces
of the chart corresponding to the probabilities.

82
Analogously to the question asked in Experiment II presented in Appendix D.
83
Alignment of preferences is important between the practitioner and the client.

160
The Steps Detailed
Starting from L = 4000 (L can vary from 4000 to 8000; the interval for L is fixed such
that it guarantees a strong preference for one lottery over the other) it was explained to
the portfolio manager that the process consists of narrowing the interval containing L
through a number of iterations until the outcome that made him indifferent between the

two lotteries is found, that is (1000, p’; 600) ∼ (L0, p’; 100).84

84
A second screen shot Display II also in Appendix B illustrates the use of the scroll bar installed to help
the decision maker understand what is required of him that is, illustrate visually that there should be a value
for which preferences between the two lotteries switched. Practitioners were savvy however and did not
need it.

161
The outcome L is thus determined through a series of iterations to make the player
indifferent between two lotteries. Each binary choice question corresponded to an
iteration in a bisection process and the indifference point is taken as the midpoint of the
last two bracketing choices as can be seen from Table 9 below which illustrates the
process for L1 and L0.0625.

Table 9: An Illustration of the Bisection Method

The Iterations Choices in Elicitation of L1 Choices in Elicitation of L0.0625

1 (-1000, 0.33;-600) vs. (-4000,0.33;-100) -6080 vs. (-7800,0.78; 0)
2 (-1000, 0.33;-600) vs. (-2500,0.33;-100) -3040 vs. (-100000,0.78; 0)
3 (-1000, 0.33;-600) vs. (-1750,0.33;-100) -4560 vs. (-100000,0.78; 0)
4 (-1000, 0.33;-600) vs. (-2100,0.33;-100) -3800 vs. (-100000,0.78; 0)
5 (-1000, 0.33;-600) vs. (-2300,0.33;-100) -4180 vs. (-100000,0.78; 0)
Indifference V. -2200 - 4370

The chosen lottery is printed in bold. The table shows that only the outcome to be elicited
is varied. Depending on the choice of the iteration, this outcome was increased or
decreased. The size of the change was always half the size of the change in the previous
question, under the restriction that the resulting outcome should be a multiple of 10 and
the resulting probability a multiple of 0.01. Otherwise, the value was set equal to the
closest multiple of 10 or of 0.01. The method resulted in an interval within which the
indifference value should lie and the midpoint of the interval was taken as the
indifference value. For example, in Table 9, the indifference value for L0.0625 should lie
between -4180 and -4560. Hence, the indifference value taken was -4370.85 The starting

85
After the final iteration, the practitioners were offered a chance to continue with the next choice or to
repeat the precedent anew, in order to to minimize the impact of response errors. Also, the number of
iterations was not the same for each step of the elicitation procedure. Indifference values were elicited in
five iterations in steps 1, 2 and 4 of the method. In step 3, seven iterations were used, the pilot sessions
undertaken by Abdellaoui et al. (2006) having shown that these numbers were sufficient to obtain the
indifference values with good precision.

162
values in the iterations were generally chosen so that the lotteries had equal expected
values. The exceptions were L1, L2, G1 and G2 whose starting values were L0 +3000, L1
+3000, G0 +3000 and G1+3000.

Once L1 is determined, it replaces L0 in the lottery and another choice question is used to
make him indifferent between the two lotteries: (L2, p’; L*) ∼ (L1, p’; L ). The probability
pl that makes the subject indifferent between L1 for sure and the prospect (L2, pl; L0), is
then determined relying thus on the probability equivalence method. The same sequence
is constructed for gains. Thus as stated above and illustrated in Table 7, in step one, two
monetary outcomes for gains and two for losses for which the difference in utility
between successive outcomes is constant, are determined in sequence as well as the two
above mentioned probabilities pg and pl for which w(pg) = w(pl ) = 0.5.

86
The following step, fractile losses, determines a sequence of eleven outcomes for
losses Lr elicited on the interval [-100000, 0] for the following utilities for losses: 0.015,
0.031, 0.062, 0.093, 0.125, 0.25, 0.375, 0.5, 0.625, 0.75, 0.875, where U (Lr) = - r using
the probability elicited for losses. Having determined pl and U( L1) set equal to −1 for
some L1 < 0 which is allowed by the uniqueness properties of the utility function in
prospect theory, the outcome L0.5 is then determined such that: (L1, pl; 0) ∼ L0.5

Immediately after, the step called Mirror Image LG determines the gain Gmax that
represents the same utility variation with respect to the 0 outcome as the absolute value of
the utility variation between the loss L0.25 and 0; this step links the utility for losses to the
utility for gains through the elicitation of three outcomes and allows the measurement of
loss aversion.

The fourth step, the fractile gains, elicits a sequence of eight outcomes between Gmax and
0 for the following utilities for gains: 0.015, 0.031, 0.062, 0.093, 0.125, 0.156, 0.187,

86
As was pointed out to me, the elicitation procedure relies on two response modes: the probability
equivalence and the fractile. The possible implication for the interpretation of the results will be discussed
in the section entitled: Concluding Remarks for Experiment I.

163
0.25, where U (Gr) = r using the probability for gains determined at the beginning. More
data was collected for losses than for gains to improve the operationalization of
Kahneman and Tversky’s (1979) definition of loss aversion. Also, many points were
elicited close to zero to operationalize Köbberling and Wakker’s (2005) definition.

The software also elicits data for two consistency tests for the measurement of utility.
The first requires that the utilities of the outcomes elicited at the beginning of the
experiment {L0, L1, L2} exhibit the same difference obtained between the successive
elements elicited in step 2 for losses after factoring out probability weighting under CPT.
Ad modem, it requires that the the utilities of the outcomes {G0, G1, G2} exhibit the same
difference as obtained between the successive elements elicited in step 4 for gains. In
other words, consistent measurement of utility requires that the equalities between utility
differences be preserved. The utility of Li and Gi , i = 0, 1, 2 were determined through
linear interpolation as shown in the next section 4.3.1. The quality of these interpolations
was however good because the Li and Gi were generally concentrated near zero where
there were many observations.

The software also allowed for a second consistency test which consists of four questions
presented to the practitioner at the end of the elicitations for gains. Each question aimed
at determining the indifference point of the paractitioner between a mixed lottery (Gr, pr;
Lr), r = 0.031, 0.0625, 0.125, 0.250, and receiving nothing that is, at determining the
probability pr that that made the negative outcomes mirror images in terms of utility of
the commensurate positive outcomes. Equation (4.2) and the results derived before show
that in each question, w + ( pr ) = w− (1 − pr ) is obtained and because w+ and w- are strictly
increasing, the same indifference probability should be found in all four questions.

164
4. 3 DATA ANALYSIS

This section shows how the shape of the utility for gains and of the utility for losses for
each subject was determined, that is, it looks at the evolution of the slope of the utility
function at various points. It also explains how the various definitions of loss aversion are
operationalized for the measurement of loss aversion, at the individual level. This section
presents first the analysis for the two consistency tests required for the non-parametric
elicitations; the data analysis for the latter is presented next, followed by the analysis for
the parametric fittings assuming power, exponential and expo-power functions. The
analysis concerning the measurement of the loss aversion coefficients in the aggregate as
well as the individual level follows.

4.3.1 Consistency Tests for Non-Parametric Elicitation

To test whether these equalities between utility differences between the successive
elements (L0, L1, L2) and (G0, G1 and G2) are preserved in the elicitation of the utility xr
for losses and for gains in steps 2 and 4 of the elicitation procedure respectively, linear
interpolation was used to calculate the utility of these Li and Gi, i = 0,1,2 labeled a0, a1,
a2 for gains and b0, b1, and b2 for losses:

U ( x2 ) − U ( x1 )
For gains U (a ) = [ ](a − x1 ) + U ( x1 ) (4.4)
( x2 − x1 )

U ( x2 ) − U ( x1 )
For losses U (b) = [ ](b − x1 ) + U ( x1 ) (4.5)
( x2 − x1 )

The x1 and x2 were chosen such as x1 < a; x2 > a; the opposite holds for b. For example, to
calculate the utility of an elicited gain a1= 2200 for a practitioner (number 7) the two
outcomes 7060 and 10580 corresponding to the utilities 0.0015 and 0.003 were chosen as
x1 and x2 respectively.

165
The trade-off method yields the same difference in utility scale and accordingly, the
hypotheses to be tested for the first consistency test are:

For gains: U (a1 ) − U (a0 ) = U (a2 ) − U (a1 )

For losses: U (b1 ) − U (b0 ) = U (b2 ) − U (b1 )

The second consistency test checks the equality of the four probabilities established for
each of the lotteries (Gr, pr; Lr), r = 0.031, 0.0625, 0.125, 0.250, towards the end of the
experiment that made the four negative outcomes mirror images in terms of utility of the
commensurate positive outcomes.

4.3.2 Non-Parametric Elicitation of the Utility Function for Gains and Losses

The shape of the utility for gains and of the utility for losses was determined by looking
at the evolution of the slope of the utility function at various points. Two elicited losses Lr
and Lr’ (for gains Gr and Gr’) are adjacent if Lr > Lr’ (Gr > Gr’) and there is no elicited

loss (gain) in between. S L↑ (r ) is defined as the slope of the segment linking (Lr, U(Lr))

and (Lr’, U(Lr’)) where Lr and Lr’ are adjacent. Similarly, S L↓ (r ) is defined as the slope of

the segment linking (Lr”, U(Lr”)) and (Lr, U(Lr)) where Lr” and Lr are adjacent. SG↑ (r ) and

SG↓ (r ) are defined analogously. ∆S L (r ) = S L↓ (r ) − S L↑ (r ) denotes the variation of the slope

around Lr when moving towards zero. Similarly, ∆SG (r ) = SG↓ (r ) − SG↑ (r ) denotes the
variation of the slope around Gr when moving away from zero. It is easily verified that
∆Si (r ), i = G, L positive, (negative, zero) corresponds to convex (concave, linear) utility.

Seven values of ∆SG (r ) were obtained for gains and eleven values of ∆S L (r ) for losses.
To account for response error, practitioners with at least four/seven negative (positive)
∆SG (r ) were classified as having concave (convex) utilities for gains. Practitioners with

at least seven/eleven negative (positive) ∆S L (r ) were classified as having concave

166
(convex) utilities for losses. Fenema and van Assen (1999) as well as Abdellaoui (2000)
and Etchart-vincent (2004) have used similar criteria.

4.3.3 Parametric Fitting of the Data

The practitioners’ data was also analyzed assuming three parametric forms: the power
family, the exponential family and the expo-power family.

The power family, characterized by a constant relative risk aversion (CRRA),87 that is,
individuals make the same proportional investments in risky assets, allows different
degrees of CRRA, which in turns allows fine tuning economic models to fit empirical data
and is used predominantly when large stakes are relevant.88 The power function is
defined as follows, with α > 0 for gains and for losses with β > 0.

x
U = U max ( )αif x ≥ 0 (4.6)
xmax
x
U = −( − ) β if x < 0
100000

The case α < 0 corresponds to a decreasing utility function. The power function for gains
is concave if α < 1, linear if α = 1 and convex if α > 1; for losses, it is convex if β <
1, linear if β = 1 and concave if β > 1.

The exponential family is characterized by a constant absolute risk aversion (CARA)

under which rich people will not be attracted to risk no matter how richer they become;

87
The risk aversion coefficients associated with the power function (CRRA) and the exponential function
form (CRRA) are shown in Table 10 in Appendix C.
88
It is conveniently used under lognormal probability distributions because the resulting risk-neutral
probability distributions which are often used to model stock prices are lognormal.

167
thus it is practical to use over small to moderate stakes. The exponential function is
defined as follows for α ≠ 0 and for β ≠ 0 .
x
U max (exp(−α ( )) − 1)
xmax
U= if x ≥ 0 (4.7)
(exp(−α ) − 1)
x
(exp(− β ( )) − 1)
U =− 100000 if x < 0
(exp(− β ) − 1)

x x
For α = 0 , it is defined as U max ( ) and for β = 0 as −
. The exponential
xmax 100000
function for gains is concave if α > 0 and convex if α < 0; for losses, it’s convex if β
>0 and concave if β < 0.89

The expo-power family was introduced by Abdellaoui, Barrios and Wakker (2002) and is
variation of a two-parameter family proposed by Saha (1993). The expo-power is defined
as follows for α ≠ 0 and for β ≠ 0 .
x
U max (1 − exp(−( )α / α ))
xmax
U= if x ≥ 0 (4.8)
1 − exp(−1/ α )
x
(1 − exp(−( ) β / β ))
U =− 100000 if x <0
1 − exp(−1/ β )

The cases where α ≤ 0 and β ≤ 0 were not considered because contrary to observation
they lead to extreme risk aversion for gains and extreme risk seeking for losses when zero
is among the outcomes of a lottery. The expo-power for gains is concave if α ≤ 1 and
convex if α ≥ 2; for losses, it is convex if β ≤ 1 and concave if β ≥ 2. obtenues pour le
mean: 5.881 E -05 et pour le median: 2.042 E -05 dans la footnote

α
89
As was pointed out to me it is and not α that should be averaged across individuals. The values
x max
obtained for the mean and median are respectively: 5.88 E-05 and 2.042 E-05.

168
4.3.4 Loss Aversion Measurement

Abdellaoui et al. (2006) was the first paper to measure loss aversion at the individual
level non-parametrically. It has done so however in the lab. This work applies their
method in the field where there is much evidence that loss aversion can explain a variety
of data. This section shows how the data obtained is analyzed in the aggregate and at the
individual level.

In the Aggregate
Loss aversion refers to the asymmetry of the value function: a steeper shape for losses
than for gains Kahneman and Tversky (1979). Accordingly, the means of the outcomes
for gains and losses corresponding to the same utilities i.e. 0.015, 0.031, 0.062, 0.093,
0.125, 0.250 are regressed linearly.

Individual Loss Aversion

There are many definitions of loss aversion. In the absence of a commonly accepted
definition, the degree of loss aversion in this work is measured according to the following
definitions: Kahneman and Tversky (1979); Wakker and Tversky (1993); Neilson (2002);
Bowman et al (1999); and Köbberling and Wakker (2005) respectively. Wakker and
Tversky (1993), Neilson (2002) and Bowman et al.’s (1999) imply both the Köbberling
and Wakker’s (2005) and the Kahneman and Tversky’s (1979) definitions.

Kahneman and Tversky (1979) have suggested that loss aversion should be defined
as −U ( − x ) > U ( x ) for x > 0 which suggests that the loss aversion coefficient could be
U ( − x)
defined as the mean or median of − over relevant x. Kahneman and Tversky
U ( x)
U (−$1)
(1992) implicitly used − as an index of loss aversion. Thus to test for loss
U ($1)
aversion in the Kahneman and Tversky (1979) sense, the loss aversion coefficient
U (−Gr )
computed for each gain amount elicited was − for all G > 0 and where r refers
U (G r )

169
to the utilities of the eight amounts of gains Gr elicited: 0.015, 0.031, 0.062, 0.093, 0.125,
0.156, 0.187, and 0.25.

First the U(–Gr) were calculated using linear interpolation for each gain elicited. Since
the number of outcomes elicited in the loss domain is greater than the number of
outcomes elicited in the gain domain, it was possible to obtain eight values of U(–Gr) for
each practitioner. When G0.25 exceeded 100000, the U(–Gr) was taken as −1 . A
practitioner was classified loss averse when at least 6 out of the 8 values of the loss
aversion coefficients computed exceeded 1, loss neutral when at least 6 values were equal
to one and not loss averse when at least 6 values were less than 1. A coefficient of loss
U (−G )
aversion was then computed for each practitioner as the mean/median of − over
U (G )
relevant G.

U ( − x) U ( y )
Neilson (2002) proposed to define loss aversion by ≥ for all positive x and y
−x y
and provided a preference foundation. A possible candidate for the coefficient of loss
U ( − x)
aversion according to this definition is the ratio of the infinum of over the
−x
U ( y)
supremum of . Hence the loss aversion coefficients were computed as the ratio
y
min(U ( Lr ) / Lr )
where the gains and losses correspond to the same utilities: 0.015,
max(U (Gr ) / Gr )
0.031, 0.062, 0.093, 0.125, 0.25. A practitioner was classified as loss averse when the
min(U ( Lr ) / Lr ) min(U (Gr ) / Gr )
ratio exceeded one and not loss averse when the ratio
max(U (Gr ) / Gr ) max(U ( Lr ) / Lr )
exceeded one.

Wakker and Tversky (1993) defined loss aversion as the requirement that U '( − x) ≥ U '( x)
for all x > 0, that is the slope of the utility function at each loss is at least as large as the
slope of the utility function at the absolutely commensurate gain and provide a preference
axiomatization. Their definition could be related to a loss aversion coefficient of the

170
 U '(− x) U '(−Gr )
mean or median of . Thus the loss aversion coefficient computed were
U '( x) U '(Gr )
where r corresponds to the following utilities: 0.015, 0.031, 0.06, 0.093, 0.125, 0.15, 0.18
and where U '(Gr ) the slope ∆ Ur/ ∆ Gr was computed as:

U '(Gr ) = 1/ 2( SG↓ (r ) + SG↑ (r ) (4.9a)

i.e.,:
(U (Gr ) − U (Gr ' )) (U (Gr " ) − U (Gr ))
U '(Gr ) = 1/ 2[ + ]
(Gr − Gr ' ) (Gr " − Gr )

with Gr’ and Gr are adjacent and with Gr and Gr” adjacent, and by defining

U '(−Gr ) = S L↑ (r ) (4.9b)

i.e.,:

(U ( Ls ) − U ( Ls ' ))
U '(−Gr ) =
( Ls − Ls ' )

if Ls’ < –Gr < Ls and Ls and Ls’ are adjacent,

U '(−Gr ) = 1/ 2( S L↓ (r ) + S L↑ (r ) (4.9c)

if –Gr = Ls. .

The slopes around each gain were computed first. Then the slopes around the
commensurate losses were computed according to the appropriate equation. Seven values
were obtained for each practitioner. A practitioner was classified as loss averse when at
least 5 out of seven values exceed one, loss neutral when six out of seven are equal to one
and not loss averse when at least six out of seven are less than one. A loss aversion
coefficient was computed for each practitioner as the mean (median) of the seven
coefficients.

A stronger definition was provided by Bowman et al (1999): loss aversion holds if

U ' ( − x ) ≥ U ' ( x ) for all positive x and y. That is the slope of the utility function for losses
is everywhere steeper than the slope of the utility function for gains. Hence, the loss

171
 min(U '( Lr )
aversion coefficient was computed as the exceeding one, and not loss averse
max U '(Gr )
min(U '(Gr )
as the exceeding one. The U '(Gr ) and U '( Lr ) were computed as in
max U '( Lr )
equations 4.9a and 4.9c for the Wakker and Tversky (1993) coefficients.

Köbberling and Wakker (2005) defined the coefficient of loss aversion as U '↑ (0) / U '↓ (0)

where the numerator and the denominator stand respectively for the left and right
derivatives of the utility at the reference point90. Hence, the loss aversion coefficient
which implies that U is steeper for small losses than for small gains was computed as the
ratio of U(L0.015)/L0.015 over U(G0.015)/G0.015 i.e. G0.015/L0.015. Subjects whose coefficients
exceeded 1 were classified as loss averse. This definition is local and exhaustive in the
sense that every practitioner could be classified as opposed to the others which are global
and where it is possible that some practitioners are left unclassified.

An implication of Köbberling and Wakker (2005)’s definition of loss aversion is that

some modeling problems are encountered when constant relative risk aversion (CRRA) is
assumed for small to modest stakes in mixed prospects but not when (CARA) that is the
exponential is assumed.

4.4 RESULTS RELATED TO THE SHAPE OF THE UTILITY FUNCTION

This section presents the results of Experiment I. The results of the two consistency tests
are presented first followed by a test for probability weighting. The results for the shapes
of the utility functions for gains and losses elicited non-parametrically are given next
followed by the parametric fittings. Whenever convenient, they are summarized in tables.
The shape of the utility function of the median practitioner is illustrated in Figure 10.

90
The ratio was informally suggested by Benartzi and Thaler (1995).

172
4.4.1 Consistency results

The paired-t tests performed show that the null hypotheses U (a1) – U (a0) = U (a2) –
U(a1) for gains and U(b1) – U(b0) = U(b2) – U(b1) for losses are not rejected (p = 0.06 for
gains and 0.65 for losses).

For the utilities 0.25, 0.125, 0.06, 0.03 the median (mean) probability values are:
0.71(0.68), 0.66(0.66), 0.71(0.66), 0.69(0.60). Friedman test shows that χ 2 = 0.881 when
µ = 3 degrees of freedom and the significance is 0.830. Paired t-tests for p0.25 and p0.125
show p = 0.559; for p0.25 and p0.06, p = 0.640; for p0.25 and p0.03, p = 0.140; for p0.125
and p0.06 , p = 0.888; for p0.125 and p0.03, p = 0.122; for p0.06 and p0.03, p = 0.041 when
45 is the degree of freedom. The individual elicited probabilities pl and pg are shown in
Table 16 in Appendix C.

4.4.2 Probability Weighting Tests

Another test was undertaken to test the equality of the probabilities for gains and for
losses. Two probabilities were elicited at the beginning of the experiment using w (p) =
0.5 for gains and losses: for the mean: p (gains) = 0.58 and p (losses) = 0.49; for the
median: p (gains) = 0.64; p (losses) = 0.46. Wilcoxon test results in p = 0.409 and paired
t-test results in p = 0.363.
The difference between p (gains) and p (losses) being far from significant, the possibility
that they are equal cannot be rejected. Because the probability equivalent method was
used to elicit these probabilities the practitioners have been rendered more risk averse and
more so for losses than for gains according to Hershey and Schoemaker (1985).

4.4.3 The Utility Function for Gains and Losses

For the shape of the utility function, the non-parametric results are given first followed by
those of the parametric fittings.

173
Non- Parametric Results
The most common pattern at the individual level was a concave utility for gains and a
convex utility for losses. Previous studies have corroborated the concavity for gains but
the evidence on the convexity of the utility function for losses has been less clear-cut.
Thus, it is important to compare the results to studies that have estimated the utility of
losses at the individual level under prospect theory following the trade-off method
(Wakker and Deneffe, 1996) and avoiding the bias due to probability weighing. The
proportion (58%) is higher than Fenema and van Assen’s (1999), Abdellaoui’s (2000)
and Etchart-Vincent’s (2004) which range between (37% and 47%) and is close to
Abdellaoui et al.’s (2006). Table 11 summarizes the results and Figure 10 shows the
shape of the utility function for the median practitioner: concave for gains and convex for
losses with a slight kink at the reference point.

Table 11: Non-Parametric Classification of the Practitioners

Losses
Concave Convex Mixed Linear Total
Concave 4 27 11 0 42
Gains Convex 0 1 3 0 4
Mixed 0 0 0 0 0
Linear 0 0 0 0 0
Total 2 28 14 0 46

174
Figure 10: The Shape of the Utility of the Median Practitioner

175
Parametric Fittings Results
The parametric estimations confirm also the concavity for gains and convexity for losses.
Table 12 summarizes the parametric fittings ( α , β ) for the three models: Power,
Exponential and Expo-Power. The individual parameters for the three models are found
in Table 16 (b) in the Appendix.

Table 12: Parametric Fittings for Utility for Gains and Losses
Power Exponential Expo-Power
Parameter α β α β α β
Mean 0.85549 1.11999 1.28799 0.64304 1.11247 1.37079
Median 0.75734 0.90049 1.01799 0.49964 1.01059 1.16821

Compared to the above mentioned studies, the median power coefficient for losses (0.9)
is within their range which varies between (0.84) and (0.97) while the median
exponential (1.01) is close to the (0.98) of Abdellaoui et al. (2006). Table 13 classifies
the practitioners according to the shape of their utilities while Table 13 (a) shows the
parametric crossing of gains by losses for the power function, the exponential and the
exponential-power respectively.

Table 13: Parametric Classification in Number and Proportion

Losses Gains
Convex 27 0.58 12 0.26
Power Concave 18 0.39 31 0.67
NA 1 0.02 3 0.06
Convex 28 0.61 12 0.26
Expo Concave 17 0.37 34 0.74
NA 1 0.02 0
Convex 18 0.39 2 0.043
Expo-
Power Concave 7 0.15 21 0.45
NA 0 0 3 0.06

176
 Table 13 (a): Parametric Crossing of Gains by Losses

Losses
Gains Power concave convex Exponential concave convex Expo-Power concave convex
concave 12 16 concave 15 18 concave 7 1
convex 0 6 convex 2 10 convex 2

4.5 RESULTS RELATED TO THE LOSS AVERSION OF PRACTITIONERS

This section presents the results for loss aversion first at the aggregate level then at the
individual level according to the definition used.

4.5.1 Loss Aversion in the Aggregate

The means for the outcomes for gains as the independent variable and losses as the
dependent for each of the utilities (0.015625, 0.03125, 0.06, 0.09375, 0.125, 0.25) were
regressed linearly through the origin. The adjusted R2 is 0.906 and the slope b is 0.450.
The lower and upper bounds of the confidence interval 0.317 and 0.583 respectively,
confirm the alternative hypothesis of b ≠ 1 i.e. reject the null hypothesis of b = 1.
b = 1 refers to a symmetrical treatment of gains and losses relative to zero. For loss
aversion to be observed the outcomes for gains need to be greater than the outcomes for
losses for the same utility in absolute value that is b < 1. In that case losses are perceived
more keenly than gains to paraphrase Kahneman and Tversky (1979). b > 1 indicates that
the satisfaction one experiences from gaining is greater than the pain associated with
losing the same amount.

4.5.2 Individual Loss Aversion

Table 14 displays the results of individual loss aversion under the various definitions and
shows as in Abdellaoui et al. (2005) that which definition is adopted matters.

177
 Table 14: Individual Loss Aversion Results under the Various Definitions

Not
Loss Loss Un-
Coefficient Averse Averse Neutral Mean Median classified
-U(-Gr)/U(Gr) 18 13 0 2.19 1.17 15
minU(Lr)/Lr/maxU(Gr)/Gr 11 6 0 1.02 0.56 29
U'(-Gr)/U'(Gr) 10 12 0 1.5 1.08 24
minU'(L)/maxU'(G) 3 2 0 0.28 0.21 41
G0.015/L0.015 17 22 3 5.95 0.74 4

According to the definition of Kahneman and Tversky (1979) 18 practitioners were found
to have at least 6 coefficients > 1 and therefore were classified as loss averse; 13 were
found to have coefficients < 1 and were classified as not loss averse. The mean (median)
was 2.19 (1.17). The percentage of loss averse practitioners (30%) is lower than
Abdellaoui et al.’s (2006) result (81%) but comparable to Bleichrodt and Pinto’s (2002)
who have used the same definition and found a range from 2.5% to 30% using the same
definition. The proportion of practitioners motivated by gains (28%) is higher than either
the latter’s proportion which varies between 0% and 2.5% or Abdellaoui et al.’s (2006)
16%.

For Neilson’s (2002) coefficient, the mean (median) was: 1.023 (0.56). Eleven
practitioners have been found to have ratios that exceed 1 and were classified as loss
averse while 6 with ratios of minG/maxL exceeding 1 were classified as not loss averse.
According to the definition of both Neilson (2002) and Kahneman and Tversky (1979),
the loss averse practitioners outnumber the not loss averse. Nevertheless and as can be
seen from the table fewer practitioners were classified according to Neilson’s (2002)
rather strict definition.

For Wakker and Tversky’s (1993) definition of loss aversion, 10 were found to be loss
averse and 12 not loss averse. The mean (median) was 1.5 (1.08) respectively. According

178
to this definition, 28% of the practitioners were found loss averse vs 37% practitioners
who were more focused on gains, as opposed to the 33% loss averse vs 24% not loss
averse of Schmidt and Traub’s (2002) who have used the same definition and the 64%
loss averse vs 12% not loss averse of Abdellaoui et al.’s (2006).

Bowman et al.’s (1999) definition of loss aversion coefficient which implies that U is
everywhere steeper for losses than for gains is clearly strict. It resulted in 3 having
min(U’(L)/max U’(G) > 1 and 2 having min(U’(G)/max U’(L) >1. The mean (median) is
0.28 (0.21).

For Köbberling and Wakker’s (2005) definition which is exhaustive in the sense that
every practitioner could be classified in one of the categories, they almost tie: 17 have a
loss aversion coefficient > 1 while 22 have a loss aversion < 1. Four were unclassified
because they had zero for the loss outcome immediately around the reference point.

Comparing the mean (median) coefficient of loss aversion found in this experiment with
Abdellaoui et al.’s (2006) and shows that the median practitioner is less loss averse than
the median student as is illustrated below in Table 15.

Table 15: Comparison of the Mean (Median) of the Practitioner and the Student
Practitioners Practitioners Students Students
Coefficient Mean Median Mean Median
-U(-Gr)/U(Gr) 2.19 1.17 2.15 1.72
minU(Lr)/Lr/maxU(Gr)/Gr 1.02 0.56 0.83 0.51
U'(-Gr)/U'(Gr) 1.5 1.08 2.02 1.53
minU'(L)/maxU'(G) 0.28 0.21 0.62 0.5
G0.015/L0.015 5.95 0.74 4.99 2.52

179
Under the definition of Kahneman and Tversky (1979), both the mean and median
values for the ratio -U(-Gr)/U(Gr) do not show consistent decrease with the size of the
gains and losses involved unlike Abdellaoui et al. (2006) findings and Bleichrodt and
Pinto’s (2002) who observed that, in the health domain, the degree of loss aversion
decreased with the size of the outcomes. Also under the definition of Wakker and
Tversky (1993) and this time similarly to Abdellaoui et al.’s (2006) this effect was not
observed. Table 16(a) immediately below show the practitioners’ individual loss aversion
Parameters.

180
Table 16 (a) : Practitioners’ Individual Loss Aversion Parameters.

Practitioner Loss Aversion Coefficients

Number K&T Neilson W &T Bowman K&W
2 2.252397 0.965014 0.074148 8.2
3 2.643696 0.011753 0.017925 0.006369
4 18.38589 0.868356 0.518233 1
5 1.063342 0.720599 1
6 5.745162 5.43307 3.12E+00 -0.00065 78.44444
7 3.651059 2.41908 3.28E+00 1 3.861842
8 0.973608 0.43299 4.16E-01 0.135577 1.913978
9 0.114753 0.146225 9.21E-02 0.03109 0.13
10 0.592003 0.36184 6.05E-01 0.271553 0.215
11 0.207838 0.105 1.43E-01 -0.00073 0.117241
12 2.143848 0.436503 1.15E+00
13 1.043437 0.314074 7.00E-01 0.211715 2.307692
14 1.923994 1.511111 1.40E+00 0.289171 5.05
15 0.955635 0.116911 1.18E+00 0.167907 0.098765
16 1.246142 0.660161 1.81E+00 0.624368 0.52505
17 4.311981 3.793549 3.05E+00 1.115939 14
18 0.646741 0.209524 0.588235
19 1.43059 0.586139 1.08E+00 0.161058 0.686792
20 0.86868 0.63938 8.27E-01 0.251432 0.512397
21 0.651103 0.289249 8.12E-01 0.268143 0.221198
22 1.330008 0.106701 2.25E+00 -0.57779 1
23 4.745734 5.44762
24 2.002662 0.222395 6.98E+00 0.270407 0.244
25 3.290824 0.320893 0.009312 0.329615
26 1.411629 0.688714 1.73E+00 0.644695 0.788599
27 0.75738 0.242836 8.23E-01 0.247444 0.236559
28 2.616173 0.246233 1.20E+00
29 0.701446 0.071101 5.67E-01 0.090685 0.088415
30 2.102039 1.258002 1.73E+00 -0.18234 2.837438
31 1.089896 0.205941 5.47E-01 0.047378 1.5
32 2.161788 1.421814 2.49E+00 0.748583 4.575
33 1.943728 1.009852 9.82E-01 0.264105 7.148936
34 2.263393 0.711339 7.35E-01
35 0.200373 0.123941 2.68E-01 0.089968 0.123563
36 5.509023 2.087996 0.039388 1.836042
37 0.888638 0.673549 7.96E-01 0.598969 1.01
38 0.539273 0.047074 0.04634 0.25
39 1.54285 0.612441 2.09E+00 0.61691 0.665323
40 0.357804 0.171257 0.150345 0.376623
41 3.711391 3.04878 2.19E+00 0.215276 50.5
42 5.733892 6.016739 3.73E+00 1.518924 50.75
43 0.802212 0.395661 0.039654 0.299517
44 0.308329 0.096127 3.41E-01 0.111308 0.131579

181
 45 1.325308 0.551901 1.46E+00 0.423769 2.962963
46 1.743035 0.693555 1.01E+00 0.284807 3.095238
47 0.788702 0.579079 9.67E-01 0.388622 0.6

As the table shows there are some blanks under the definitions of Wakker and Tversky
(1993), those of Bowman et al.’s (1999) and those of Köbberling and Wakker (2005). For
min(U '( Lr )
Bowman et al.’s (1999) definition which is restrictive, calculating involves
max U '(Gr )

u ( xr ) − u ( x r −1 )
initially calculating a ratio of differences equal to where r is the reference
x r − x r −1

utility and varies accordingly. The blanks correspond to xr − xr −1 equal to zero, that is
equal outcomes for successive utilities. They belong to practitioners who have the
numbers: 5, 12, 18, 23, 28, 34. The first one # 5 has equal outcomes for r = 5, 6, # 12 has
also equal outcomes for r = 5 and 6 while those numbered 12, 23, 28, 34 have equal
outcomes for r = 1 and 2.

For Wakker and Tversky (1993) definition, the blanks correspond to the practitioners
whose elicited outcomes were identical for two consecutive utilities. The practitioners’
numbers for these are: 2, 5, 18, 23, 38, 40, 43 mostly for r whose utility is 0.125, 0.156,
while those numbered 3, 4, 25, 36 have huge coefficients.

For Köbberling and Wakker’s (2005) definition, those numbered 12, 23, 28, 34 have zero
for the first loss outcome.

Because the findings for the shape of the utility function show different curvatures for the
utility of gains and the utility of losses, one may expect that the global measures of the
utility function, which measure loss aversion at different points, do not indicate loss
aversion unambiguously. A comparison of the practitioners loss aversion coefficients
across definitions in Table 16(a) shows that there are big differences between the global
measures of Wakker and Tversky’s (1993) and Kahneman and Tversky’s (1979)

182
coefficients on one hand and the local coefficient of Köbberling and Wakker’s (2005) on
the other hand. Bowman et al.’s (1999) and Neilson’s (2002) are global measures but
have turned out to be too restrictive for empirical purposes. The results hence, argue in
favor of a separation of loss aversion from the curvature of the utility function.

183
CONCLUDING REMARKS FOR EXPERIMENT I

Part I has shown extensively, that prospect theory can explain a variety of field and
experimental data that are paradoxes under expected utility. Abdellaoui et al.’s (2006)
method allows a test of the theory. Applied in the financial field, the method proved to be
unproblematic. The availability and the ease of the application of a method to elicit
prospect theory’s utility on the whole domain is important for the theory’s application
and test.

An advantage of Abdellaoui et al.’s (2006) procedure which is based on the elicitation of

utility midpoints is the control over the endpoints (in Table 9, L2 is specified). Vind’s
(2003) is a similar method which elicits utility midpoints for given endpoints but requires
more measurements, three to one per utility midpoint. It also assumes instead of prospect
theory a general additive representation which is not directly applicable to prospect
theory.

The elicitation procedure, nevertheless, uses a mix of response modes. First, in step one
the probability equivalence method is used to elicit a probability which is then, used to
elicit outcomes through the fractile method. According to Hershey and Shoemaker (1985)
the two methods of utility assessment are not equivalent and the order in which they are
presented, matters. For both gains and losses, the authors predict that for the above
mentioned order the practitioners would relatively be less risk averse in the second mode

(p: 1222). Put differently, if p’ is elicited such that G1 ∼ (G2, p’; G0) from a subject, then

by asking the same subject to state G such that G∼ (G2, p’; G0), the G obtained is < G1.
Thus, it is possible that the probability of gain has been adjusted upwards in the first
mode showing an increase in risk aversion and the sure amount adjusted downward in the
second mode (p: 1216). Also, to reduce order effects, counterbalancing the elicitation
steps for gains and losses across practitioners might reduce the possible exaggeration or
the dampening of the curvature of the utility function by certain features of the elicitation
procedure. Most of the practitioners have indeed shown some aspects of fatigue by the

184
time they finished the fractile losses. It did look that losses were tiring to consider and
the fractile for losses long.

It remains to be said regarding the applicability of the method that the utility graph
immediately obtained at the end of the computerized session motivated the financial
practitioners to take time to discuss the psychological elements of CPT and most of them
asked for a summary of the experiment’s results to be sent to them. The reason is that
they perceived CPT as reflecting the major aspects of their decision making process
which could be summarized in an expression often used by them and quoted by Thaler
(1993, p: 513): “small profits and large losses.”

Regarding the results per se, the elicitation of the utility function in Experiment I shows
that the most common pattern for the utility function for financial practitioners is
concavity for gains and convexity for losses. The proportion (58%) is higher than Fenema
and van Assen’s (1999), Abdellaoui’s (2000) and Etchart-Vincent’s (2004) which range
between (37% and 47%) and is close to the (54%) of Abdellaoui et al.’s (2006). For
parametric fittings, the median power coefficient for losses (0.9) is within their range
which varies between (0.84) and (0.97).

Loss aversion exists in the aggregate (b = 0.450, adjusted R2 = 0.906). At the individual
level the coefficients of loss aversion vary however with the definitions used emphasizing
the need for a commonly accepted definition. Few practitioners could be characterized
according to the global definitions of Neilson (2002) and Bowman et al.’s (1999) which
seem overly strong for empirical purposes. Contrasted with the results of previous studies
that have estimated a loss aversion coefficient according to the same definition, the
findings show that for Wakker and Tversky’s (1993) definition, the practitioner’s median
is found lower than both Schmidt and Traub’s (2002) and Abdellaoui et al.’s (2006) i.e.
1.08 vs 1.43 vs 1.53 respectively. For Kahneman and Tversky’s (1979) definition it is
close to Abdellaoui et al.’s (2006), lower than Bleichrodt et al.’s (2001) i.e. 1.69 vs 1.72
vs 2.17 respectively. Finally, for the local definition of Köbberling et Wakker (2005) the
median portfolio manager is rather not loss averse relative to the median student of

185
Abdellaoui et al. (2006): 0.74 vs 2.52 respectively. In brief, as Table 15 which compares
the practitioners’ and students’ means (medians) shows the former is less loss averse than
the latter.

Professionals who are exposed to a range of training and high level of knowledge differ
in their assessment of the stakes from the students. The difference in the degree of loss
aversion may be due to the practitioners’ range of training and high level of knowledge
relative to the students.

Also, the interviews were conducted during the period (2003-2004) which corresponds to
a growth (the Standard and Poor index was up by 28.17% and 10.9% respectively relative
to the preceding year). It is not inconceivable thus that the degree of loss aversion of the
practitioners diminished during that period with the upward movement of the stock
market. Barberis, Huang and Santos’ (2001) asset pricing model is based on changing
risk aversion over time generated by introducing loss aversion over financial wealth
fluctuations and allowing the degree of loss aversion to be affected by prior investment
performance.91 Without the variation of loss aversion with past movements in the stock
market, the authors couldn’t account for the high volatility of stock returns observed.

In addition to original prospect theory (1979) their model is influenced by the

psychological findings in Thaler and Johnson (1990) and Gertner (1993). The latter two
studies examine the effect of a sequence of gains and losses for small stakes and high
stakes respectively and show evidence of house money effect92 i.e., prior gains cushion
following losses rendering the individual less loss averse.

The two studies also show that the controverse is true, i.e., the individual is more loss
averse after having incurred losses. The finding implies that the convexity of the utility

91
Barberis, Huang et Santos (2001) show that time-varying loss aversion degree is necessary alongside loss
aversion to account for the equity premium puzzle.
92
The terminology “playing with the house money” refers to the gamblers increased willingness to bet
when ahead.

186
function can not be due to the integration of sequential gambles93 but does not refute the
risk seeking in the domain of losses. Indeed, another result of Thaler and Johnson (1990)
is that students are risk seeking when the lottery results in consequences which offer the
chance of a breakeven and limit the maximal loss.

Gross (1982) well-documents the get even phenomenon in the stock market which he
refers to as the “getevenitis disease”. The remedy is painful: to realize losses and perceive
sunk costs as irrelevant to the decision making, as dictated by rationality, practitioners
need to give up the hope they might get even before they get out. 94

Last but not least, an important incentive for the practitioner to seek gain is the
compensation package he is eligible to in case of profits. For instance, for the record
profits of the year 2006, the “Wall St Bonus” according to the New York Times (Dec. 25,
2006)95 varied from $1mil to $3 mil for an average managing director to $60 mil for the
investment houses like Goldman Sacks, Lehman Brothers and Morgan Stanley and more
for a select group of hedge fund managers and private equity executives.

The variation of loss aversion with the volatility of the market (5% only of the
practitioners do beat the market, the majority gains in a bull market and loses in a bear
market) and the luring compensation package practitioners could get in case of profits are
factors to be considered. Nevertheless, the utility function must reflect the preferences of
the individual and the utility must not change with the method being used. Hence, the
preferences of the students/practitioners analyzed following a different method must
reflect consistent preferences. Experiment II presented next investigates non-
parametrically albeit qualitatively the preferences of MBA students, potential financial

93
If investors did integrate many years of stock market gains and losses, they would essentially be valuing
absolute levels of wealth and not the changes in wealth that are so important to prospect theory (Barberis,
Huang and Santos, 2001).
94
Daniel Kahneman and Jonathan Renshon recently argued in Foreign Policy magazine that the American
administration’s unwillingness to face reality in Irak reflects a basic human aversion to cutting one’s losses,
the same instinct that makes the gambles stay at the table hoping to break even (Krugman, P., The New
York Times, Jan, 8, 2007)
95
For instance, the bonus awarded to Blankfein, L.C. CEO of Goldman Sacks is $54.3 mil (New York
Times, Dec 25, 2006).

187
practitioners and tests for loss aversion using stochastic dominance criteria newly
developed by Baucells and Heukamp (2006).

188
CHAPTER V. EXPERIMENT II.

LABORATORY INVESTIGATION USING STOCHASTIC DOMINANCE

CONDITIONS

Just as Abdellaoui et al.’s (2006) elicitation method contains all the essentials of Wakker
and Deneffe (1996) but not vice versa, Baucells and Heukamp’s (2006) newly developed
conditions contain all the essentials of Levy and Levy’s (2002) but not vice versa. That is
their conditions include Levy and Levy (20002) and other properties. In that sense, the
newly developed conditions also represent a case of all of the lower is in the higher but
not all of the higher is in the lower as was pointed out first by Aristotle96. To recapitulate
in brief, preferences can be inferred from choices among lotteries constructed such as one
stochastically dominates the other. An individual with a concave non-decreasing utility
function i.e. a risk averse expected utility maximizer, will dis-prefer a lottery that is
dominated by SSD and the converse is also true: if the individual eliminates SSD
dominated alternatives, he has a concave utility function.

To discriminate between classes of non-decreasing utility functions which are not

concave throughout, other conditions are needed. Ad initium, Moshe Levy and Haim
Levy (2002) develop Prospect Stochastic Dominance (PSD) and Markowitz Stochastic
Dominance (MSD) to differentiate between utility functions which are concave for gains
and convex for losses (have an S shape as postulated by Kahneman and Tversky (1992))
and those which are convex for gains and concave for losses (have an inverse S as
postulated by Markowizt (1952)).97 Their experimental investigation (Levy and Levy,
2002) which uses these conditions rejects the S shape and supports the inverse S.

96
Aristotle first pointed out that the impulse of evolution is this not vice versa which invariably establishes
a hierarchy, an increasing order of wholeness. For instance, a molecule includes atoms yet has properties
that are not merely the sum of its atoms. To spot the higher from the lower in any sequence, Wilber (1996,
p: 28) suggests a thought experiment where all of the higher is destroyed and none of the lower.
97
The utility function as defined by Kahneman and Tversky (1979, p: 279) is the exact opposite of
Markowitz’ (1952) postulated utility function as shown by Figure 11 illustrated further below.

189
It turns out however, as Baucells and Heukamp’s (2004) and Wakker’s (2003) show98
that the reason the S shape is rejected is that Levy and Levy (2002) have assumed that
the probability weighting function is unlikely to play a role for probabilities ≥ 0.25. Had
they not, the results would have been compatible with both theories.99

To be able to discriminate empirically between the two theories, Baucells and Heukamp
develop stochastic dominance conditions which incorporate probability weighting and
loss aversion. Their experimental investigation using these conditions (Baucells and
Heukamp, 2006) “rules out” the possibility that the utility function is concave throughout,
or has an inverse S shape and confirm that the S shape holds assuming the curvature of
the utility function is the same on each side of the real line. However, loss aversion is
only evident when the probability for gains/losses is similar or the same in the two
lotteries. In brief, for mixed gambles and for the purpose of prediction: the overall
probability matters and loss aversion competes with the convexity of the utility function
for losses (Baucells and Heukamp, 2006).

Experiment II, investigates the risk preferences of MBA students at ASU using the
Baucells and Heukamp (2006) lottery pairs. More specifically, the students were asked to
answer a questionnaire of 20 tasks and to choose for each between two lotteries
constructed as head to head competition between the CPT and Markowitz’ (1952)
theories.

Baucells and Heukamp (2006) having hypothesized the impact of the overall probability
in mixed gambles, Experiment II has also included tests of this effect. The questionnaire
format allowing it, the students were asked to comment on the choices they have made.
The statements either confirmed or disturbed the motivation inferred for a particular

98
Wakker (2003, p: 981) and Baucells and Heukamp (2004) independently and using different methods
show that the results are compatible with CPT’s predictions contrary to Levy and Levy ’s (2002) claims.
99
For lottery F in Levy and Levy’s (2002) experiment 2, all 4 outcomes have p = 0.25 and w(0.25) = 0.29
close to 0.25; but the correct decision weights for the four outcomes which depend on cumulative
probabilities, are respectively: 0.29, 0.16,0.13 and 0.29, showing the decision weights for the extreme
outcomes to be twice as much as those for the intermediate outcomes (Wakker, 2003).

190
choice particularly when a switching from one theory to the other occurs. The method
however does not quantify loss aversion.

Chapter V consists of six sections. Section 1 gives the intuition for prospect and
Markowitz stochastic dominance followed by the characterization of preferences using
these conditions. Section 2 states the objective of the experiment and describes the source
of the data. The analysis of the data is presented in section 3. The results related to the
shape of the value function are presented next, followed by the results related to loss
aversion and the results related to the overall probability of gain /loss respectively. A
final section concludes.

5.1 PROSPECT AND MARKOWITZ STOCHASTIC DOMINANCE

The insight in brief behind Levy and Levy’s (2002) prospect dominance condition (PSD)
and Markowitz dominance condition (MSD) is that a prospect theory follower gives more
importance to outcomes near the origin than to extreme outcomes while the opposite is
true for a Markowitz follower assuming no probability weighting. However, when
probability weighting is factored in the decision making process more importance is
given to extreme outcomes relative to intermediate outcomes. Hence, for CPT not to
remain ambiguous, it’s necessary to generalize the PSD, MSD conditions to capture this
important aspect of CPT and also the other remaining aspect of the theory namely loss
aversion. This is the motivation for the Baucells and Heukamp’s (2006) newly developed
SD conditions. Section 5.1.1 presents first Levy and Levy’s (2002) conditions to be
followed by Baucells and Heukamp’s (2006) in section 5.1.2.

5.1.1 Levy and Levy’s (2002) Stochastic Dominance Criteria

This section presents for clarification the insight that led Levy and Levy (2002) to claim
that prospect theory is “much ado about nothing”. It will be followed with the preference
characterization.

191
The Intuition
If prospect stochastic dominance (Markowitz stochastic dominance), that is PSD (MSD),

is to be used analogously to SSD then, a lottery F preferred or equivalent to G (F ≽ G)

according to PSD (MSD) ought to necessarily imply that the expected utility of F is
greater than the expected utility of G for any agent with an S shape (inverse S shape)
utility function. Assuming no probability weighting the difference in expected utility is
given by100:
b
∆ = ∫ [G(t ) − F (t )]U '(t )dt (5.1)
a

Where F and G are the cumulative distributions of lotteries F and G and where it is

assumed for simplicity that both F and G take the values 0 and 1 for some a ≤ 0 and b ≥
0 respectively. a and b correspond to the two extreme inflection points in the Markowitz
utility function and are expected to be at extreme wealth levels. Figure 11 contrasts the
latter utility function with its exact opposite prospect utility function pointing at the same
time to a commonality between the two: in both, decision makers base their decisions on
changes in wealth relative to some reference point.

Figure 11: Markowitz and Prospect Theory Utility Function

100
Equation (5.1) is obtained by integrating by parts the difference in expected utility of the two lotteries F
and G (Rothshild and Stiglitz, 1970).

192
According to equation (5.1) the difference [G(t) –F(t)] is scaled in proportion to U’(t),
which is to say that segments where the slope of U(t) is high are stretched more relative
to segment where it is small. The areas near the origin are magnified for a prospect theory
follower since the slope is higher near the origin while the extremes are magnified for a
Markowizt’ follower since this is where the slope is higher.

The prospect stochastic dominance and Markowitz stochastic dominance conditions

developed by Levy and Levy (2002) capture this insight and characterize the preferences
of a decision maker who maximizes the expectation of the utility function assuming no
probability weighting. Depending on whether the decisions are based on total wealth or
changes in wealth101, PSD and MSD cover EU or any reference model of which however
w(p) is not a part.

Levy and Levy’s (2002) Preference Characterization

The Levy and Levy (2002) conditions characterize both the S shape and the inverse S
shape utility functions. Individuals however, may have different types of preferences as
can be seen from the results of Experiment I: alongside the S shape utility function
characterizing the majority of practitioners, a concave utility function for gains and losses
for instance is also representative of the practitioners’ preferences. The S shape and the
inverse S shape utility functions are considered then as two classes of preferences which
are subsets of a general utility function which describes individuals who prefer more to
less. This function is monotonic and non-decreasing characterized by a first derivative
which is never negative. Formally:
U ∈ U1 if U ' ≥ 0 .
The preference class of non-decreasing concave preference functions Uconcave is
characterized in addition by a non-increasing second derivative. Formally:
U ∈ Uconcave if U ' ≥ 0 and U " ≤ 0 .
While the class of non-decreasing convex preference functions Uconvex is characterized
by a non-decreasing second derivative:

101
Levy and Levy (2002, p: 1338) show that FSD, SSD, PSD MSD can be stated in terms of total wealth or
changes in wealth. Levy (1992) is a review article of stochastic dominance rules.

193
U ∈ Uconvex if U’≥ 0 and U”≥ 0.

These classes are subsets of U1 as are also the two utility functions S shape (inverse S
shape) illustrated in Figure 11 since they are all non-decreasing functions ( U ' ≥ 0 ).
Formally, with UP, the set of S shape utility functions,
U ∈ UP if U ' ≥ 0 for all x ≠ 0, U " ≥ 0 for x < 0 and U " ≤ 0 for x > 0,
and with UM, the set of inverse S- shape utility functions,
U ∈ UM if U ' ≥ 0 for all x ≠ 0, U " ≥ 0 for x > 0 and U " ≤ 0 for x < 0.

Proposition 5.1:
Define F and G as above. Then, F fPSD G for all S-shape utility functions, U ∈ UP, if and
only if

0
∫ y
[G (t ) − F (t )]dt ≥ 0 for all y ≤ 0 and
x
∫ [G(t ) − F (t )]dt ≥ 0 for all x ≥ 0 hold
0

where there is a strict inequality for some pair (y0, x0) and for some U0 ∈ UP

Proposition 5.2:
Define F and G as above. Then, F fMSD G for all inverse S-shape utility functions, U ∈
UM if and only if

y
∫ [G(t ) − F (t )]dt ≥ 0 for all
a
y ≤0 and
b
∫ [G(t ) − F (t )]dt ≥ 0 for all
x
x ≥ 0 hold

(with at least one strict inequality)

Markowitz stochastic dominance rule is not necessarily the opposite of prospect

stochastic dominance rule. However, they are opposites if the two distributions have the
same mean. Thus, for F and G with equal means, F fPSD G iff G fMSD F .

194
These conditions were used by Levy and Levy (2002) as a guide in designing pairs of
lotteries or tasks that are head to head competitions between the two alternative theories.
The choices of the decision maker between the so-designed lotteries reveal properties of
his preferences and their representations under a choice model. Thus, for a given model,
the prospect stochastic dominance (Markowitz stochastic dominance) condition related to
a specific feature of the model allows for testing the qualitative properties of the latter.

For two lotteries F and G designed such as F fPSD G and G fMSD F the choice of F (G)
implies that the utility function of the decision maker has an S shape (inverse S shape).
Figure 12 illustrates the cumulative distributions for the two lotteries F and G of Task I102
designed such as F fPSD G and G fMSD F . The solid lines represent F and the dashed lines
G. The signs correspond to the signs of [G(t) –F(t)]. Because F(t) < G(t) in the areas
which are closer to the origin, these will be magnified for a prospect theory follower.

Figure 12: The Cumulative Distributions for Task I: F fPSD G and G fMSD F .

1.00

0.75
Cum. Prob.

0.50

0.25

0
-8000 -6000 -4000 -2000 0 2000 4000 6000

Outcomes
102
All tasks are described in the questionnaire in Appendix D.

195
However, as argued above, Levy and Levy (2002) have assumed that the probability
weighting function is unlikely to play a role for probabilities ≥ 0.25 motivating Baucells
and Heukamp (2006) to develop stochastic dominance criteria that account for the
probability weighting function. Their preference conditions are presented next preceded
by the insight behind them for clarification.

5.1.2 Baucells and Heukamp’s (2006) Stochastic Dominance Criteria

To be able to discriminate between prospect theory (1992) and Markowitz’ utility theory,
Baucells and Heukamp (2006) extend the stochastic dominance conditions to CPT by
incorporating the probability weighting function and loss aversion. The intuition behind
the extension is presented first followed by the preference characterization.

The Intuition
The difference in expected utility was given in equation (5.1) assuming no probability
weighting. However, when probability weighting of extreme events is factored into
equation (5.1), the difference in utility between lotteries F and G is given by:
0 b
∆ = ∫ [ w(G(t )) − w( F (t ))]U '(t )dt + ∫ [ w(1 − F (t )) − w(1 − G(t ))]U '(t )dt (5.2)
a 0

Accordingly, in parallel of the horizontal stretching, the vertical cumulative probability

axis (0, 1) is stretched by the probability weighting function making the prospects near 0
(possibility) and near 1 (certainty) more desirable and magnifying the difference between
F(t) and G(t) in these areas. Thus, while F fPSD G for a prospect theory follower
equation (5.2) can still yield a preference for G. To resolve the ambiguity, stochastic
dominance conditions that incorporate the probability weighting function need to be used
in constructing the lotteries.

196
Accounting for the Probability Weighting Function
The qualitative features that a descriptively relevant probability weighting function (pwf)
exhibits are concavity for low values of p (p close to 0) and convexity for high values of
p (p close to 1). To capture these features, five classes of pwfs are defined: W0 is the class
of convex probability weighting functions and W1 the class of concave pwfs, Wc is the
class of pwfs convex between c and 1, Wd the class of pwfs concave between 0 and d and
Wcd = Wc ∩ W d is their linear intersection (w(p) = p) which contains segments that are

convex between c and 1 and segments that are concave between 0 and d. The c and d
denote both c+ and c- (or both d+ and d-) which apply to w+ and w- respectively.
Wcd = Wc ∩ W d . If 0 < c ≤ d < 1, then Wcd is necessarily linear and agrees with the

Kahneman and Tversky (1992) inverse S shape pwf: “shallow in the middle interval and
changes abruptly towards the ends of the probability interval (0,1). Figure 13 illustrates
the Wcd class:

197
 Figure 13: the Wcd class of probability weighting function

The classes of probability weighting function thus defined, the areas where the vertical
stretching and the horizontal stretching are compounding each other that is, the intervals
of the payoff line in which the curvatures of u and w are conjugate103 are delineated
restricting the pwf to Wcd . First order stochastic dominance conditions are imposed then
in the intervals where the vertical stretching runs against the horizontal stretching that is
in the intervals where the utility function and the pwf do not have conjugate curvatures in
order to extend the SD conditions to CPT. Figure 14 shows for c ≤ 1/6, the cumulative

103
In the loss domain, a convex u and a convex w- are conjugate as are a concave u and a concave w-. In the
gain domain, a concave u and a convex w+ are conjugate as are a convex u and a concave w+.

198
distributions for the two lotteries F and G of Task VII designed such as F fPWSD G and

G fMWSD F .

Figure 14: The Cumulative Distributions for Task VII: F fPWSD G and G fMWSD F .

1.
1-c
d
0.75
Cum. Prob.

0.50

0.25
1-d
c
0
-8000 -6000 -4000 -2000 0 2000 4000 6000

Outcomes

Formally, Baucells and Heukamp’s (2006) stochastic dominance conditions which

account for the probability weighting are stated as follows:

Proposition 5.3:
− +
F fPWSD G iff F ≽ G for all U ∈ UP , w− ∈ Wcd− , and w+ ∈ Wcd+ . Similarly,
− +
F fMWSD G iff F ≽ G for all U ∈ UM , w− ∈ Wcd− and w + ∈ Wcd+ ;

− +
F fSWSD G iff F ≽ G for all U ∈ Uconcave , w− ∈ Wcd− , and w+ ∈ Wcd+ ; and
− +
F fS *WSD G iff F ≽ G for all U ∈ Uconvexe , w − ∈ Wcd− , and w+ ∈ Wcd+

199
Accordingly, to predict preferences between two lotteries when neither stochastically
dominates the other, the probability weighting function must be restricted. The choice of
c and d is ultimately the result of a tradeoff. When c = 1 and/or d = 0, the four conditions
are reduced to first order stochastic dominance. Thus, by decreasing c or increasing d,
first order stochastic dominance is imposed on a smaller range which increases the scope
of application of the different stochastic dominance conditions. However, the Wcd might
become too narrow and might not contain the desired functions. Their choice of c and d
was selected from a range of [0.05 to 0.88] according to the sensitivity analysis they
undertook.

Accounting for Loss Aversion

To incorporate loss aversion, Wakker and Tversky’s (1993) definition is used to define
the class of utility functions possessing loss aversion. Since the latter guarantees that the
stretching of the horizontal axis at − x is at least as large as the stretching of the
horizontal axis at x > 0. This allows to use positive first order stochastic dominance
segments in the losses domain where F ( − x) < G ( − x) are used to counteract negative first
order stochastic dominance segments in the gains domain where G (x) < F(x). Loss
aversion entails comparisons between the positive and the negative domain hence the
sign-dependent pwfs need to be constrained. Assuming c+ < d+ and c- < d- the slope of
the linear segment of w+ is defined as s+ and the slope of the linear segment of w- is
defined as s- and the condition s − ≥ s + is imposed. The slope of w- being larger than the
slope of w+ ensures that areas of positive FSD in the negative domain can counteract
areas of negative FSD in the positive domain. Defining UL as the class of utility
functions such that:
U ∈ UL if U '( − x) ≥ U '( x) for all x ≥ 0,

Baucells and Heukamp’s (2006) stochastic dominance condition which accounts for loss
aversion is stated formally as follows:

200
Proposition 5.4:
− +
F fPLSD G iff F ≽ G for all U ∈ UL , w − ∈ Wcd− and w + ∈ Wcd+ such that s − ≥ s +

The probability weighting functions for positive outcomes and those for negative
outcomes are both restricted since under CPT the decision weights for the former are
calculated independently from the decision weights for the latter. The condition s − ≥ s + is
consistent with the empirical finding that w- exhibits less deformation than w+ (Tversky
and Kahneman, 1992; Abdellaoui, 2000). Furthermore, it reinforces loss aversion since it
implies that the stretching of the vertical axis for negative outcomes is larger than the
vertical stretching for the positive outcomes. Figure 15 illustrates the cumulative
distributions of Task XIX designed such as to test for loss aversion such as F fPW d LSD G
c

0.65
for W0.15 .

Figure 15: The cumulative distributions for Task XIX

1.00

0.75
Cum. Prob.

0.50

0.25

0
-6000 -4000 -2000 0 2000 4000 6000

Outcomes

201
Finally, with UPL = UP ∩ UL, the full CPT SD condition which accounts for probability
weighting and for loss aversion is stated formally as follows:

Proposition 5.5:
− +
F fPWLSD G iff F ≽ G for all U ∈ UPL , w − ∈ Wcd− and w+ ∈ Wcd+ such that s − ≥ s +

The extension of the SD conditions to CPT allows for testing hypotheses about the
curvature of the utility function and/or loss aversion assuming the pwf is inverse S
shaped. It also allows to test the curvature of the latter assuming the empirical
specifications for CPT for the utility function hold. Finally, if one assumes all the
empirical specifications for CPT holds, then a violation of the SD condition implies a
violation of the CPT model.

5.2 EXPERIMENTAL APPLICATION

The tasks constructed by Baucells and Heukamp (2006) using the newly developed
stochastic dominance conditions were as in Levy and Levy (2002) head to head
competition between two prospects F and G having the same mean: F f G for the S-
shape functions and G f F for the inverse S. Therefore, the preferences of the subjects
can be inferred from their choices of F or G. For two lotteries F and G designed such
as F fPWSD G and G fMWSD F , as in Task VII for instance with c = 1/6 and d = 2/3, the
choice of F implies that the hypothesis of U ∈ UP is not rejected i.e. is consistent with an
S shape utility function and that the hypothesis of U ∈ UM is rejected.

Lotteries with at least three outcomes are required because the design of the tasks using
the stochastic dominance conditions implies always the addition on common extreme
outcomes. Experiment II uses these tasks to investigate the risk preferences of MBA
students. Section 5.2.1 states the objectives of Experiment II, section 5.2.2 describes the
source of the data.

202
5.2.1 Objective of the Experiment

Experiment II has three objectives: the first is to investigate what type of utility function
MBA students have, i.e., is it S-shape as in CPT or inverse S as in Markowitz; the second
is to find out whether they are loss averse or not and the third is to shed light on the effect
of the overall probability of gain.

5.2.2 Source of Data

The 40 subjects were first and second year MBA students at Arizona State University,
(Phoenix-Arizona, U.S.). They were contacted through the student services coordinator
associate following the recommendation of William Boyes, Professor of economics at
ASU. The interviews were conducted in groups of 4-5 in the suites, private rooms, the
MBA students at the W.P. Carey School of Business at ASU have access to for team
work. The students were working and/or had worked previously. (One requirement for
MBA at ASU is a minimum of three years of experience). They were paid $10 for the 30
mn on average needed to complete the questionnaire.

The questionnaire consisted of 22 tasks. For 20 of these, the students had to choose
between two investments F and G introduced as in Baucells and Heukamp (2006) and
Levy and Levy (2002) that is as follows: “Suppose that you decided to invest $10000
either in stock F or Stock G. Which stock would you choose, F or G when it’s given that
the dollar gain or loss one month from now will be as follows.”

The students were also asked to state the reason for their choice. For the last 2 tasks as in
(Payne, 2005) students were asked whether they preferred to increase the overall
probability for gains when given a chance to. The questionnaire handed out to the
students is in Appendix D104.

104
The questionnaire is based on the paper Baucells, M. presented at the FUR (2004) at GRID, Cachan,
France which he has co-authored with Heukamp H. and which I attended.

203
5.3 ANALYSIS OF DATA

The 20 tasks used in this experiment were designed initially in response to Levy and
Levy’s (2002) claim that prospect theory is much ado about nothing. The basis of the
authors’ claim is that subjects who were given a choice in three tasks I-III, where
F fPSD G and G fMSD F opted for G the Markowitz dominating lottery in the following
proportions: 71%, 62% and 76% respectively. To refute Levy and Levy’s (2002) claim
Baucells and Heukamp (2006) use the tasks I-III as they are initially designed (assuming
no probability weighting) to ensure similar responses from subjects then, construct
modifications of these tasks which incorporate the probability weighing function.

Thus, tasks I-III mimic Levy and Levy’s (2002) tasks to ensure that U ∉ UP where UP is
the set of S-shape utility functions. Because F fPSD G if w is linear (convex) throughout,

then the choices favoring G imply that U ∉ UP .

In contrast to tasks I-III where no probability weighting is assumed, tasks IV through

VIII, discriminate between the hypotheses of an S shape utility function or an inverse S
shape, assuming the empirical specification of the probability weighting function holds.
These tasks are modifications of tasks I-III and exhibit F fPWSD G and G fMWSD F with

w ∈ W0.1
0.74
, for d ≤ 0.74 and c ≥ 0.1. F is the expected answer if U ∈ UP and G is the
expected answer if U ∈ UM. Thus, the choice of F in these tasks i.e. U ∈ UP, given the
results of Tasks I-III, implies that the probability weighting function can be neither linear
nor convex throughout.

To specifically examine the curvature of the latter function ( pwf ) and its change near the
origin, task IX was designed as a modification of task I with a slight change in the
probability attached to common outcomes. It exhibits F fPWSD G and G fMWSD F and

204
tests conjointly with task I whether w ∈ W0.02 and whether w ∈ W0 can be rejected
respectively.

Before however, testing for loss aversion, four tasks are designed to refute the argument
that the concavity for gains and the convexity for losses are driven by the certainty effect.
Thus, tasks X-XI are constructed as all gains gambles, and test the concavity for gains
and tasks XII- XIII are all losses gambles and test the convexity for losses. In all tasks,
none of the outcomes is certain and all satisfy the F fPWSD G for c ≥ 0.1.

An important feature of CPT is loss aversion and seven tasks are designed to test for it.
The tasks XIV-XX which test for loss aversion are divided into two categories. Tasks
XIV-XVI have F fW d LSD G and test solely loss aversion with no assumption on the utility
c

function. For these tasks the choices favoring F imply then loss aversion. The remaining
tasks XVII- XX test a joint hypothesis of loss aversion and concavity for gains/convexity
for losses and exhibit all F fPW d LSD G and can be interpreted as a global test for CPT. For
c

tasks XVII and XX, the choices favoring F imply that the subjects are loss averse and
that the shape of their utility function is concave for gains, while for tasks XVIII and
XIX, choices favoring F imply that the subjects are loss averse and that the shape of their
utility function is convex for losses.

Finally, Baucells and Heukamp (2006) having hypothesized the impact of the overall
probability in mixed gambles, the last two tasks were included to investigate the effect of
the overall probability of winning or losing. The students were presented with the
following mixed lottery designed by Payne (2005): ($100, 0.2; $50, 0.2; $0, 0.2; $-25,
0.2; $-50, 0.2) and they were asked in a first step to add ($38) to either $0 or $100 and in
a second step to add ($38) to either $50 or $100. The choice of adding the ($38) to the
($0) would imply a preference to increase the overall probability of gain.

205
5.4 RESULTS RELATED TO THE SHAPE OF THE UTILITY FUNCTION

The results related to the shape of the utility function are summarized in Table 17 below
which also compares them to those obtained by Baucells and Heukamp (2006).

Table 17: Results of Experiment II Regarding the Shape of the Utility Function in %
Task I II III IV V VI VII VIII IX X XI XII XIII
B.& H. 36 42 33 61 84 66 76 84 81 79 74 75 64
ASU 37.5 40 37.5 67.5 85 87.5 85 92.5 92.5 85 87.5 72.5 85

For tasks I-III which mimic the Levy and Levy (2002) lotteries, the proportions of
students who chose F over G are 37.5 %, 40 %, and 37.5% respectively that is, results
comparable to Levy and Levy’s (2002) are obtained .

For tasks IV-XIII designed to investigate the utility function in mixed lotteries, the high
majority of the students chose F. Regarding the relatively low proportion (67%) for task
IV, which is complex and presents identical probabilities, the reasons the students gave
were mostly centered on the middle class outcomes, which agrees with the editing of
extreme outcomes that have the same probabilities. However, in task VI, which also
presents common outcomes with common probabilities, the higher overall probability of
gain invoked by most students increased the proportion to 87%.

For task IX which is a slight modification of task I (in F a maximum amount has been
added with a probability of 2% and in G a minimum amount has been added also with a
probability of 2%) the major reversal of preference from G for task I to F for task IX
(37.5% to 92.5%) shows that w ∈ W0.02 is a plausible class of pwfs and emphasizes the

206
abrupt change of the pwf near the origin ( w ∉ W0 ). It also suggests that decision makers
use the range of outcomes for as a decision criterion.

For tasks X-XIII, the responses are overwhelmingly in favor of a higher p of gain and
minimizing extreme loss. The p-values are significant and reject the null hypothesis that
% F = 0.5 for tasks IV- XIII. The p-values for binomial for all tasks are in Table 20.

5.5 RESULTS RELATED TO LOSS AVERSION

The results pertaining to tasks XIV-XVI designed to test solely for loss aversion are
summarized in Table 18 below which also compares them to the results obtained by
Baucells and Heukamp (2006).

Table 18: Results for Tasks XIV-XVI in %

Task XIV XV XVI
B.&H. 43 61 64
ASU 60 67.5 55

For task XIV, 60% of the students chose F to avoid a loss. The majority of the 40% who
picked G said they were not attracted to the 80% probability of no gain in F. For this task
the students at ASU show more loss aversion than Baucells and Heukamp’s (2006)
subjects (60% vs 43%) respectively. However, the latter’s subjects were composed of
students and professionals who for the same task have chosen differently: the students’
percentage of F compared to the professionals’ was (48% vs 35%) for task XIV. That is
students show more loss aversion than professionals for losses in Baucells and
Heukamp’s (2006). To recall, a comparison of the results of Experiment I with those of
Abdellaoui et al. (2006) has also shown that the median student is more loss averse than
the median practitioner.

207
For task XV where the stakes are higher (maximum loss has increased) than in XIV the
proportion of students who chose F increased (67.5%) pointing to an increase in loss
aversion in mixed lotteries as the size of outcomes increases. Students also invoked the
decrease in the probability of no gain to 60% going from XIV to XV. In Baucells and
Heukamp (2006), both the proportions of students and professionals who chose F
increased (65% vs 54% respectively).

For task XVI which examines loss aversion closer to the origin relative to task XV, less
students (55 %) preferred F to G preferring the higher p for the higher gain in G for the
same overall probabilities of gain/loss, the overall probability of loss having increased
from XIV to XV to XVI.

Tasks XVII-XX present a global test for CPT. Except for task XVII, which shows a
preference for the higher overall probability of gain, F is clearly preferred to G in line
with the predictions of CPT. They are summarized in Table 19 below:

Table 19: Results for Tasks XVII-XX in %

Task XVII XVIII XIX XX
B. & H. 46 70 74 77
ASU 40 85 70 87.5

For task XVII as in Baucells and Heukamp, 60% of the students chose G for the higher
probability of gain (70% in G vs 50% in F), or the higher probability for the higher gain
and in their own words “$500 is not too much to lose”. For task XVIII and XIX 85%,
70% respectively of the students chose F which necessitates a convex utility function for
losses. For task XX, 87.5% chose F which necessitates a concave utility function.

208
5.6 RESULTS RELATED TO THE OVERALL PROBABILITY OF GAIN/LOSS
For the last two tasks, 82% vs 67.5% respectively of the students confirmed that the
overall probability of gain/loss is an important factor in deciding between two
investments. Table 20 summarizes all the results for Experiment II and shows their p-
values.

Table 20: Results for all Tasks and their P-Values for Binomial

Task B&H% ASU % p-values

I 36 37.5 0.113846
II 42 40 0.205903
III 33 37.5 0.113846
IV 61 67.5 0.026857
V 84 85 9.55E-06
VI 66 87.5 2.10E-06
VII 76 85 9.55E-06
VIII 84 92.5 7.62E-08
IX 81 92.5 7.62E-08
X 79 85 9.55E-06
XI 74 87.5 2.10E-06
XII 75 72.5 0.004427
XIII 64 85 9.55E-06
XIV 43 60 0.205903
XV 61 67.5 0.026857
XVI 64 55 0.527089
XVII 46 40 0.205903
XVIII 70 85 9.55E-06
XIX 74 70 0.011412
XX 77 87.5 2.10E-06
XXI 85 9.55E-06
XXII 67.5 0.026857

209
CONCLUSION

It has been recognized since their introduction by Rothschild and Stiglitz (1970) that
second order stochastic dominance criteria offer a framework to test different features of

the theory under which they apply. To recall from section 3.4.1, if F ≽ G according to
second order stochastic dominance then necessarily that implies that the expected utility
of F is greater than the expected utility of G for any agent with a non-decreasing concave
utility function. Recently, Levy and Levy (2002) have designed and used stochastic
dominance criteria that apply within EU or any reference dependent model that does not
however assume probability weighting an important source of departure from expected
utility maximization.

Baucells and Heukmap (2006) have thus designed stochastic dominance conditions that
apply within CPT, the theory which, being developed as a descriptive alternative to EU
has been quite successful at explaining its violations. When these conditions are used as a
guide to design pairs of lotteries, the choices of the decision maker which reveal his
preferences and their representation under CPT allow a non-parametrical test of the
qualitative properties of CPT.

The result is that joint hypothesis on the curvature of the utility function and the
probability weighting functions can be tested in three ways: 1) If one assumes the CPT’s
empirical specifications for the probability weighting function hold, then the
corresponding stochastic dominance conditions can be used to test hypotheses about the
curvature of the utility function and/or about loss aversion; 2) if one assumes the CPT’s
empirical specifications for the utility function hold, then a test of the hypothesis on the
shape of the probability weighting function can be undertaken and 3) if one assumes all
the CPT’s empirical specifications hold, then a violation of the corresponding stochastic
dominance condition implies a violation of the CPT model.

210
Experiment II undertaken using the Baucells and Heukmap’s lotteries tests these joint
hypotheses on 40 MBA students at ASU. The results show the following:

1) assuming the CPT’s empirical specifications for the probability weighting function
hold, the results of tasks XI, XIII and VII show that the hypothesis of U ∈ UP was not
rejected i.e., is consistent with an S shape U while the hypotheses of U ∈ UM and U ∈
Uconvex, U ∈ UM and U ∈ Uconcave, and U ∈ UM were rejected respectively. Also loss
aversion’s results tested in tasks XIV, XIV and XVI show that the hypothesis of U ∈ UL
was not rejected.

2) assuming the CPT’s empirical specifications for the utility function hold, the curvature
of the probability weighting function was tested in Tasks I and IX, the result of which,
the hypothesis that w ∈ W0 was rejected but not the hypothesis that w ∈ W0.02.

3) assuming all the CPT’s empirical specifications hold, consistency with CPT was tested
in tasks XIV and XVII-XX. Except for task XVII where the hypothesis U ∈ UPL is
rejected the results show consistency with CPT. The results of tasks XXI and XXII
provide evidence however, that the overall probability of gain matter and task XVII is a
case where students have reported that their choice for G (and hence the shift from loss
aversion to gain seeking) was motivated by the higher overall probability of gain (70%)
and/or the high probability of the maximal gain while the loss in not extreme.

In brief, the results of Experiment II as their (2006)’s reject the hypothesis that the utility
function is convex for gains and concave for losses and are consistent with an S-shape
utility function. In addition, they show consistency with loss aversion provided the
probabilities are similar or the same and last but not least they confirm the importance of
accounting for the probability weighing function.

211
GENERAL CONCLUSION

Baucells and Heukamp (2006) have generalized and extended second order stochastic
dominance conditions for expected utility to prospect theory. A new method has been
added thus for the experimentalists who are interested in falsifying a particular hypothesis
about the shape of the utility function or about the probability weighing function without
having to elicit these functions.

For those experimentalists interested however in exploring the entire utility function
under prospect theory Abdellaoui et al. (2006) offer a complete parameter-free elicitation
procedure.

The availability of the two methods motivated this experimental work which aims at
inferring individuals’ preferences from the choices presented to them and which consists
of two experiments. Experiment I applies Abdellaoui et al’s (2006) method in the field to
elicit completely and non-parametrically the utility functions of financial practitioners
and to measure their individual loss aversion degrees without however committing to any
particular definition of loss aversion. Experiment II infers the preferences of MBA
students using Baucells and Heukamp’s method (2006).

Generally, the standard preference function presupposed in elicitation procedures results

from expected utility the normative model of choice which describes how rational agents
ought to choose. Nevertheless, for a normative model to be operational and prescriptively
useful the actual behavior of an individual in simple choice settings must be compatible
with the behavior assumed in the model. Expected utility has been found to be violated
systematically in experimental works to the extent it is not defensible as a descriptive
model of actual behavior as was shown in chapter II. For the selection of a theory which
might show, when tested experimentally a good approximation of the underlying
preference function the individual is assumed to optimize, Part I has reviewed the models
presupposing a single preference function, an important tenet of coherence, yet defensible

212
as a descriptive models of actual behavior. It was shown that prospect theory explains the
anomalies and the basic phenomena EU is used to explain by integrating psychological
insights into economics. These include the reflection effect, loss aversion and the
subjective treatment of probabilities (the section “support from experimental findings” in
the general introduction relates the evidence at a glance). In brief, prospect theory has
been successful at organizing empirical departures from EU maximization due to a
combination of empirical realism and theoretical advantages.

The availability of a method to elicit preferences under prospect theory is thus important
for practical interests. Prospect theory has refined the understanding of risk aversion to a
large extent. Under CPT risk attitude three components that are affected jointly by a
gamble: a reference dependent utility function, loss aversion and a probability weighting
function challenging as a consequence the role of utility in representing solely risk
attitude. The notions of risk aversion defined behaviorally i.e. defined independently of
any model are not equivalent anymore to the concavity of the utility function and the
elicited shape of the latter implies different risk attitudes depending on which model
assumed. The equivalence, for instance, of the convexity of the utility function to risk
seeking under EU does not hold under prospect theory because of the non-linearity of the
probability function in the latter105.

Moreover, a parameter-free method that does not depend on the appropriateness of the
selected function has several advantages: 1) it could validate reasonable functional forms
and thus the use of certain types of parametric estimation procedures. These have the
advantage of smoothing response errors while relatively good estimates can be obtained
with a smaller set of lotteries; 2) non-parametric measurements give insights into the
psychological reasoning that underlies the data because elicited utilities can be directly
traced back to observed choices; and 3) they give empirical meaning to the concepts

105
For example, if a subject indicates that he is indifferent between a sure loss of $40 and the two-outcome
1
prospect ( −$100, ;$0), then equation (4.1) reveals that this risk seeking preference is consistent with a
2
concave utility for money if w− ( )<0.4.
1
2

213
underlying the decision theory and hence are particularly useful to prescriptive decision
analysis.

Experiment I which elicits the utility function of financial practitioners under prospect
theory following Abdellaoui et al.’s (2006) corroborate their findings that the most
common pattern is concavity for gains and convexity for losses. A comparison however
of the individual loss aversion degrees with Abdellaoui et al.’s (2006) results for students
shows that the median practitioner is less loss averse than the median student.

The results of Experiment II which infers the preferences of MBA students using
Baucells and Heukamp (2006) turn out to be consistent with an S-shape utility function
and with loss aversion provided the probabilities are similar or the same. Hence, the
preferences of the students are consistent across methods. A noteworthy result of
Experiment II is that the majority of the students (60%) who have shifted from being loss
aversion to being not loss averse in task XVII invoked either the higher overall
probability of gain or higher probability of maximal gain combined with a limited
extreme loss. The result hints to more than one mechanism of risk attitude being affected
jointly by a gamble and that the behavior loss averse/not loss averse depends on which
mechanism106(s) is or are triggered.

Professionals who are exposed to a range of training and high level of knowledge and are
offered a powerful incentive son a yearly basis, differ in their assessment of the stakes
from the students. Also, the interviews were conducted during the period (2003-2004)
which corresponds to a growth (the Standard and Poor index was up by 28.17% and
10.9% respectively relative to the preceding year). It is not inconceivable thus that the
degree of loss aversion of the practitioners diminished during that period with the upward
movement of the stock market (Barberis, Huang and Santos, 2001).

106
Mechanisms are frequently occurring and easily recognizable causal patterns that are triggered under
generally unknown conditions or with indeterminate consequence. Type (B) obtains when two causal
chains are triggered leaving the net effect indeterminate, type (A) obtains when one cannot predict which of
the causal chains will be triggered (Elster, J., 1999)

214
Reflecting on what could induce risk seeking in the domain of losses, some aspects of the
practitioners’ work discussed during the interviews come to mind. In order to realize
losses as dictated by rationality, practitioners need to give up the hope they might “get
even” before they get out. This is an occupational aspect that MBA curriculums may find
it advantageous to emphasize and to offer some hard rules against. The explicit
formulation of decision problems in terms of final assets to eliminate risk seeking is
precisely the one advice Daniel Kahneman chose to give to financial practitioners on
CNBC the American channel for financial news upon winning the Nobel Prize.107

Finally, what emerged from the discussions with practitioners is that they find it difficult
to admit having been wrong to their peers and clients. Being evaluated according to their
performance in a very competitive environment which depends on astuteness in
judgment, it seems to be a “hard pill to swallow” (Kleinfield, 1983). According to them,
the soundness of a judgment looms very large in the real that is uncertain environment
they work in, because it is the starting point of what they call the investment chain
process. Indeed, a major concern of money managers is the prediction of the direction of
macro events (trade deficits, inflation/deflation…)108 with which begins the investment
process, i.e. their judgments of the probability of occurrence of uncertain events.
Ultimately, it seems this is what distinguishes the performance of one practitioner from
another independently of the market’s ups and downs. Among the portfolio managers
interviewed some were able to predict correctly the probability of the 2001 market crash
and “went short” sustaining an above average performance during that period.109
Accordingly, CPT seems to describe a natural thought-pattern in a general population in
which the different components of risk attitude are quite mixed and addresses the

107
Amos Tversky who has died was ineligible for the prize because the Nobel is not awarded
posthumously.
108
The investment process as a chain was described as follows by one hedge fund manager: Macro→
Sector→ Company→ Sustainability →Alignment. The analysts are responsible for the sector and the
company while the manager covers the macro, the sustainability of the fund and the alignment of its
objectives with the clients’.
109
Under risk with probability weighting filtered, these few have in common in addition to their sound
judgment, a convex utility for gains and an almost linear utility for losses.

215
practitioner’s concerns in a way the normative EU does not. Yet theories that describe the
behavior of individuals exhibiting extreme predominance of one component such as EU
and/or DT are useful because any extreme position is more uncompromisingly clear and
therefore more easily recognized and understood than the intermediate positions which
do not in any way contain or reconcile the extreme positions (Huxley, 1945). In that
sense, a theory describing individuals who exhibit extreme predominance of loss aversion
may be helpful.

216
REMARKS AND DIRECTIONS FOR FUTURE RESEARCH

Combined, the results of Experiment I and those of Experiment II point to the importance
of reference dependence as part of economic modeling as advocated by Rabin (1996)
among others.

The convexity of the utility function in the domain of losses is an effect of a reference-
dependence. More precisely, it is a psychological framing effect that results from shifting
reference points. What is required then of the practitioner who refuses to take losses
perceived as such relative to his reference point is to bring about a stronger and opposite
affect a la Spinoza110 by explicitly formulating his decision problem in terms of final
wealth. The required task is demanding and points, pending the results of further
research on reference-indexed preferences, to the benefits of using corrective models111 in
decision analysis to help the practitioners make better decisions for their own interest as
well as the interest of their clients.

Loss aversion is another effect of reference-dependence. Shalev (2000) describes the

existence of this phenomenon as the most striking result of reference-dependent utility
functions and extends the analysis of games which models interactive choice such as the
behavior in markets to include both reference dependence and loss aversion. The first
step however to test his prediction that different degrees of loss aversion would lead to
different equilibrium strategies in game theory is to measure loss aversion at the
individual level.

110
The 16th century philosopher has, according to Damasio (2003) intuited in Ethics the new findings in
neuro-science hundreds of years ago: Feelings are foundational components of the mind and a negative
affect cannot be neutralized except by a positive and stronger affect brought about by reasoning and
intellectual effort.
111
As argued by Bleichrodt, Pinto and Wakker’s (2001).

217
The task so far has been problematic. Kahneman et al.’s (1991) estimates of the observed
disparities between the willingness to pay (WTP) and the willingness to accept (WTA),
show the median and mean WTA values to be between 1.4 and 16.5 times as large as the
corresponding WTP values. The comparison however may have been affected by factors
such as substitution and income effects. The problem encountered in Experiment II is the
absence of an agreed-upon definition of loss aversion which emphasizes the need of a
precise meaning for the latter.

Perhaps, to be more uncompromisingly clear and therefore more easily recognized, loss
aversion which is formalized in prospect theory needs to be disentangled from the two
other components of risk aversion in the theory, namely the curvature of utility for gains
and losses and probability weighting. The separation also essential for research on
varying reference points112 is however difficult theoretically and empirically.

Empirically, the interpretation of utility as independent of other factors and prior to risk is
so far controversial (a summary of the debates is given in Abdellaoui, Barrios and
Wakker, 2003) and there is no independent empirical implication yet for probability
weighting although attempts have been made at establishing a psychological rationale for
the probability weighting function (Weber, 1994; Gonzalez and Wu, 1999; Wakker,
2003).

Theoretically, Schmidt and Zank (2005) argue that the equivalence of the behavior to a
utility steeper for losses than for gains under the original prospect theory, does not hold
under CPT unless the two weighting functions for gains and losses are convex and

112
Bleichrodt, Pinto and Wakker (2001) observe the loss aversion index by comparing the kink of the
utility function at a point when it is the reference point with the kink at the same point when another point
is the reference point. They do so however assuming basic utility is the same for different reference points
although in a reference dependent model, the utility is determined not only by the outcome but also by the
relationship of the outcome to the reference point.

218
 113
coincide. Their characterization (2001) of risk aversion in CPT through a joint
condition on utility curvature, probability weighting and loss aversion supports however
the Köbberling and Wakker’s (2005) index which arises naturally in their framework.

Köbberling and Wakker (2005) consider loss aversion as a logically independent

component of risk attitude. Nevertheless, their work is based on the assumption that the
utility function is smooth at the zero point allowing the different processing of gains and
losses to be captured by the kink.

The axiomatic foundation of Köbberling and Wakker’s (2005) degree of loss aversion
without these restrictions awaits perhaps the observation of certain features that have yet
to reveal themselves. In response to what a friend of Einstein has told him once in a
jest:114

“The mathematician can do a lot of things, but never what you want him to do just at the
moment.”

Einstein explained that the theorist’s work falls in two tasks of entirely different nature.
He must first discover certain features and then use his skills to draw the conclusions
which follow from them.115 Amos Tversky and Daniel Kahneman have showed that these
features which are not allowed to destroy each other may dwell in more than one

113
Schmidt and Zank (2001) argue that for the definition of loss aversion expressed in terms of the
properties underlying utility function to be useful in decision analysis ad instar the Rothschild and Stiglitz’
(1970) definition of risk aversion, it must have also the same behavioral implications in different theories.
114
Einstein quoted his friend in his Inaugural Address to the Russian Academy of Sciences in 1914 (Barnes
and Noble Books, 1934).
115
A case in point is that before Quiggin (1982, p: 328) observed that “while individuals may distort the
probability of an extreme outcome in some way, they need not treat intermediate outcomes with the same
probability in the same fashion” there was nothing he could do to formalize the observation.

219
discipline116 and Daniel Kahneman has won for that insight the Nobel prize an award that
honors the most prize-worthy discovery in a year’s nominations.

116
Camerer, Lowenstein and Prelec (2003) point in the direction of neuroscience as potential candidate, if
not directly then indirectly, through its impact on psychology.

220
Bibliography

Abdellaoui, M. 2000. Parameter-Free Elicitation of Utility and Probability Weighting

Functions. Management Science 46 1497-1512.

Abdellaoui, M. 2002. A Genuine Rank-Dependent Generalization of the von-Newmann-

Morgenstern Expected Utility Theorem. Econometrica 70 717-736.

Abdellaoui, M., B. Munier. 1998. The Risk-Structure Dependence Effect: Experimenting

with an Eye to Decision Aiding. Annals of Operations Research 80 237-257.

Abdellaoui, M., H. Bleichrodt, C. Paraschiv. 2006. Measuring Loss Aversion under

Prospect Theory: A Parameter-Free Approach. Working Paper, GRID, ENSAM-
ESTP, Paris, France and Erasmus University, Rotterdam, The Netherlands.

Abdellaoui, M., C. Barrios, P.P. Wakker. forthcoming. Reconciling Introspective Utility

with Revealed Preference: Experimental Arguments Based on Non-Expected
Utility. Journal of Econometrics.

Abdellaoui, M., F. Vossmann, M. Weber. 2005. Choice-Based Elicitation and

Decomposition of Decision Weights for Gains and Losses under Uncertainty.
Management Science 51 1384-1399.

Allais, M. 1953. Le Comportement de L’Homme Rationnel devant le Risque: Critique

des Postulats et Axiomes de L’Ecole Américaine. Econometrica 21 503-546.

Allais, M. 1979. The So-Called Allais Paradox and Rational Decisions under
Uncertainty. M. Allais and O. Hagen, eds. Expected Utility Hypothesis and the
Allais Paradox. Reidel, Boston/Dordrecht, 437-681.

Anscombe, F., R. Aumann. 1963. A Definition of Subjective Probability. Annals of

Mathematical Statistics 34 199-205.

Arkes, H. 1991. Costs and Benefits of Judgment Errors: Implications for Debiasing.
Psychological Bulletin 110 486-498.

Arrow, K. 1965. Aspects of the Theory of Risk-Bearing. Yrjo Jahnsson Saatio, Helsinki.

221
Arrow, K. 1995. Interview with Kenneth Arrow. The Region. Federal Reserve Bank of
Minneapolis, U.S.

Barberis, N., M. Huang, T. Santos. 2001. Prospect Theory and Asset Prices. Quarterly
Journal of Economics 66 1-53.

Barnes, J., J. Reinmuth. 1976. Comparing Imputed and Actual Utility Functions in a
Competitive Bidding Setting. Decision Sciences 7 801-812.

Bateman, I., Munro, A., Rhodes, B., C. Starmer, R. Sugden. 1997. A Test of the Theory
of Reference-Dependent Preferences. Quarterly Journal of Economics 62 479-
506.

Baucells M., H. Heukamp. 2004. Reevaluation of the Results of Levy and Levy (2002a).
Organizational Behavior and Human Decision Processes. 94 15-21.

Baucells M., H. Heukamp. 2006. Stochastic dominance and Cumulative prospect Theory.
Management Science 52 1409-1423.

Baumol, J. 1958. The Cardinal Utility which is Ordinalist. Economic Journal 68 665-672.

Bell, D. 1982. Regret in Decision Making under Uncertainty. Operations Research 30

961-981.

Benartzi, S., R. Thaler. 1997. Myopic Loss Aversion and the Equity Premium Puzzle. D.
Kahneman and A. Tversky, eds. Choices, Values and Frames. Cambridge
University Press, Cambridge, 301-316.

Bentham, J. 1789. Introduction to the Principles of Morals and Legislation. Ed. 1948
Blackwell, Oxford.

Bentham, J. 1824. The Book of Fallacies. J. Peregrine Bingham, H.L. Hunt, eds. London.

Bernoulli, D. 1738. Specimen Theoriae Novae de Mensura Sortis. Commentarii

Academiae Scientarium Petroplitanae 5 175-192. Translated into English, Louise
Sommer, 1954. Exposition of a New Theory on the Measurement of Risk.
Econometrica 22 23-36.

Birnbaum, M. 1974. The Nonadditivity of Personality Impressions. Journal of

Experimental Psychology 102 543-561.

222
Blaug M. 1980. The Methodology of Economics. Cambridge University Press,
Cambridge.

Bleichrodt, H., J. Pinto. 2000. A Parameter-free Elicitation of the Probability Weighting

Function in Medical Decision Analysis. Management Science 46 1485-1496.

Bleichrodt, H., J. Pinto. 2002. Loss Aversion and Scale Compatibility in Two-Attribute
Trade-Offs. Journal of Mathematical Psychology 46 315-337.

Bleichrodt, H., J. Pinto, J., P.P. Wakker. 2001. Making descriptive Use of Prospect
Theory to Improve the Prescriptive Use of Expected Utility. Management Science
47 1498-1514.

Bowman, D., D. Minehart, M. Rabin. 1999. Loss Aversion in a Consumption-Savings

Model. Journal of Economic Behavior and Organization 38 155-178.

Breiter, H. et al. 2001. Functional Imaging of Neural Responses to Expectancy and

Experience of Monetary Gains and Losses. Neuron 30 619-639.

Budescu, D. V., W. Weiss. 1987. Reflection of Transitive and Intransitive Preferences, a

Test of Prospect Theory. Organizational Behavior and Human Decision
Processes 39 184-202.

Burks, A.W. 1977. Chance, Reason, Cause: An Inquiry into the Nature of Scientific
Evidence. The University of Chicago Press, Chicago.

Cachon, G., C. Camerer. 1996. Loss Avoidance and Forward Induction in Coordination
Games. Quarterly Journal of Economics 112 165-194

Camerer et al. 1997. Labor Supply of New York City Cabdrivers: One Day at a time. The
Quarterly Journal of Economics 112 407-41.

Camerer, C. 1988. Prospect Theory in the Wild: Evidence from the Field. D. Kahneman
and A. Tversky, eds. Choices, Values and Frames, Cambridge University Press,
Cambridge, 288-300.

Camerer, C. 1989. An Experimental Test of Generalized Utility Theories. Journal of Risk

and Uncertainty 2 61-104.

223
Camerer, C. 1992. Predictions about Indifference Curves in the Unit Triangle: A Test of
Variants of Expected Utility Theory. Journal of Economic Behavior and
Organization 18 391- 404.

Camerer, C., T.-H. Ho. 1994. Violations of the Betweenness Axiom and Nonlinearity in
Probability. Journal of Risk and Uncertainty 8 167-196.

Camerer, C., M. Weber. 1992. Recent Developments in Modeling Preferences:

Uncertainty and Ambiguity. Journal of Risk and Uncertainty 5 325-370.

Camerer, C., G. Lowenstein, D. Prelec. 2003. Neuroeconomics: How Neuroscience Can

Inform Economics. Working Paper, California Institute of Technology, U.S.

Carbone, E., J.D. Hey. 1995. Stochastic Choice with Deterministic Preferences: An
Experimental Investigation. Economics Letters 47 161-167.

Chateauneuf, A., P.P. Wakker. 1999. An Axiomatization of Cumulative Prospect Theory

for Decision under Risk. Journal of Risk and Uncertainty 18 137-145.

Chateauneuf, A., M. Cohen. 1994. Risk Seeking with Diminishing Marginal Utility in a
Non-Expected Utility Model. Journal of Risk and Uncertainty 9 77-79.

Chew, S-H., K. MacCrimmon. 1979. Alpha-nu Choice Theory: A Generalization of

Expected Utility Theory. Working Paper (669) University of British Columbia.

Chew, S-H., K. MacCrimmon. 1979. Alpha Utility Theory, Lottery Composition and the
Allais Paradox. Working Paper (686) University of British Columbia.

Chew, S-H., W. Waller. 1986. Empirical Tests of Weighted Utility Theory. Journal of
Mathematical Psychology 30 55-72.

Chew, S-H., L. Epstein, L., U. Segal. 1991. Mixture Symmetry and Quadratic Utility.
Econometrica 59 139-163.

Chew, S-H., E. Karni, Z. Safra. 1987. Risk Aversion in the Theory of Expected Utility
with Rank-Dependent Probabilities. Journal of Economic Theory 42 370-381.

Cohen, M., J.-Y. Jaffray, T. Said. 1987. Experimental Comparisons for Individual
Behavior under Risk and under Uncertainty for Gains and for Losses.
Organizational Behavior and Human Decision Processes 31 1-22.

224
Cramer, G. 1728. Letter of G. Cramer to N. Bernoulli. Translated into English, Louise
Sommer. Bernoulli, D. 1954. Exposition of a New Theory on the Measurement of
Risk. Econometrica 22 23-36.

Damasio, A. 2003. Looking for Spinoza: Joy, Sorrow and the Feeling Brain. Harcourt
Brace, U.S.

Damasio, A. et al. 2000. Subcortical and Cortical Brain Activity during the Feeling of
Self-Generated Emotions. Nature Neuroscience 3 1049-1056.

de Finetti, B. 1937. La Prévision, ses Lois, ses Sources Subjectives. Ann. Inst. H.
Poincaré 7 1-68. Translated into English, H.E. Kyburg, Jr. 1964. Foresight: its
Logical Laws, its Subjective Sources. H.E. Kyburg and H.E. Smokler, eds.
Studies in Subjective Probability. Wiley, New York, 2nd ed. 1980, Krieger, New
York.

Diamond, W. D. 1988. The Effect of Probability and Consequence Levels on the Focus
of Consumer Judgments in Risky Situations. Journal of Consumer Research 5
280-283.

Diecidue, E., P.P. Wakker. 2001. On the Intuition of Rank-Dependent Utility. The
Journal of Risk and Uncertainty 23 281-298.

Edgeworth, F. Y. 1881 Mathematical Psychics: An Essay on the Application of

Mathematics to the Moral Sciences. ed. 1961. Augustus M. Kelley, New York.

Edwards, K. 1996. Prospect Theory: A Literature Review. International Review

Financial Analysis 5 18-38.

Edwards, W. 1955. The Prediction of Decisions among Bets. Journal of Experimental

Psychology 50 201-214.

Eeckhoudt, L., C. Gollier. 1992. Les Risques Financiers: Evaluation, Gestion et Partage.
Ediscience, Paris.

Einstein, A. 1914. Inaugural Address to the Prussian Academy of Science. Barnes and
Noble. eds. Albert Einstein: Essays in Science. Barnes and Noble, U.S.

225
Ellsberg, D. 1961. Risk, Ambiguity and the Savage Axioms. Quarterly Journal of
Economics 74 643-669.

Elster, J. 1999. Alchemies of the Mind: Rationality and the Emotions. Cambridge
University Press, Cambridge.

Epstein. L., S. Zin. 1989. Substitution, Risk Aversion, and the Temporal Behavior of
Consumption and Asset Returns: A Theoretical Framework. Econometrica 57
937-969.

Etchart-Vincent, N. 2004. Is Probability Weighting Sensitive to the Magnitude of

Consequences? An Experimental Investigation on Losses. Journal of Risk and
Uncertainty 28 217-235.

Farquhar P. 1984. Utility Assessment Methods. Management Science 30 1283-1300.

Fenema, H., M. van Assen 1999. Measuring the Utility of Losses by Means of the Trade-
off Method. Journal of Risk and Uncertainty 17 277-295.

Fiegenbaum A., H. Thomas. 1988. Attitudes towards Risk and the Risk Return Paradox:
Prospect Theory Explanations. Academy of Management Journal 31 85-106.

Fischer, G., M. Kamlet, S. Fienberg, D. Schkade. 1984. Risk Preferences for Gains and
Losses in Multiple Objective Decision Making. Management Science 32 1065-
1086.

Fischsoff B. 1982. Debiasing. D. Kahneman, P. Slovic and A. Tversky, eds. Judgment

under Uncertainty: Heuristics and Biases. Cambridge University Press,
Cambridge, U.K.

Fischoff B. 1991. Value Elicitation: Is there Anything There? American Psychologist 46

835-847.

Fishburn, P. 1967. Methods of Estimating Additive Utilities. Management Science 13

435-453.

Fishburn, P. 1970. Utility Theory for Decision Making. Wiley, New York.

Fishburn, P. 1982. The Foundations of Expected Utility. Reidel, Dordrecht, The

Netherlands.

226
Fishburn, P. 1988. Nonlinear Preference and Utility Theory. John Hopkins University
Press, Batimore, U.S.

Fishburn, P. 1989. Retrospective on the Utility Theory of von Neumann and

Morgenstern. Journal of Risk and Uncertainty 2 127-158.

Fishburn, P., G. Kochenberger. 1979. Two-Piece von Neumann-Morgenstern Utility

Functions. Decisions Sciences 10 503-518.

Fox, C., A. Tversky. 1995. Weighting Risk and Uncertainty. Pychological Review 102
269-283.

Fox, C., B. Rogers, A. Tversky. 1996. Options Traders Exhibit Subadditive Decision
Weights. Journal of Risk and Uncertainty 13 5-17.

Friedman, M., L.J. Savage. 1948. The Utility of Choices Involving Risk. The Journal of
Political Economy 56 279-305.

Gayant, J.P. 2004. Rôle de la Transformation des Probabilités dans la Combinaison

d’Actifs Risqués. Annales d’Economie et de Statistiques 73 141-155.

Georgescu-Roegen, N. 1936. The Pure Theory of Consumer’s Behavior. Quarterly

Journal of Economics 50 545-593.

Gertner, R. 1993. Games Shows and Economic Behavior: Risk Taking on Card Sharks.
Quarterly Journal of Economics 150 507-521.

Gilboa, I., D. Schmeidler. 1989. Maximin Expected Utility with Non-Unique Prior.
Journal of Mathematical Economics 18 141-153.

Gilboa, I. 1987. Expected Utility with Purely Subjective Non-Additive Probabilities.

Journal of Mathematical Economics 16 65-88.

Glick, I. 1957. A Social Psychological Study of Futures Trading. Ph.D. Dissertation,

University of Chicago.

Gneezy, U., J. Potters. 1997. An Experiment on Risk Taking and Evaluation Period.
Quarterly Journal of Economics 112 631-645.

227
Gonzalez, R., G. Wu. 1999. On the Shape of the Probability Weighting Function.
Cognitive Psychology 38 129-166.

Grauer, R. 1981. A Comparison of Growth Optimal and Mean Variance Investment

Policies. Journal of Financial and Quantitative Analysis 16 1-21.

Grayson, C. J. 1960. Decisions under Uncertainty: Drilling Decisions by Oil and Gas
Operators. Graduate School of Business, Harvard University, Cambridge,
Massachusetts.

Green, J., B. Jullien. 1988. Ordinal Independence in Nonlinear Utility Theory. Journal
of Risk and Uncertainty 1 355-387.

Green, P.E. 1963. Risk Attitudes and Chemical Investments Decisions. Chemical
Engineering Progress 59 35-40.

Grether, D., C. Plott. 1979. Economic Theory of Choice and the Preference Reversal
Phenomenon. American Economic Review 69 623-638.

Gross, L. 1982. The art of Selling Intangibles: How to Make your Millions ($) by
Investing Other People’s Money. New York Institute of Finance, New York.

Gul, F. 1991. A Theory of Disappointment in Decision Making under Uncertainty.

Econometrica 59 667-686.

Hadar, J., W. Russell. 1969. Rules for Ordering Uncertain Prospects. American Economic
Review 59 25-34.

Halter, A. N., and G.W. Dean. 1971. Decisions under Uncertainty. South Western
Publishing Cincinnati, U.S.

Hammond, P. 1988. Consequentialist Foundation for Expected Utility. Theory and

Decision 25 25-78.

Handa, J. 1977. Risk, Probabilities and a New Theory of Cardinal Utility. Journal of
Political Economy 85 97-122.

Hanoch, G., H. Levy. 1969. The Efficiency Analysis of Choices Involving Risk. Review
of Economic Studies 36 335-346.

228
Hardie, B.J.S., E. J. Johnson, P.S. Fader. 1993. Modeling Loss Aversion and Reference
Dependence Effects on Brand Choice. Marketing Science 12 378- 393.

Harless, D. W., C. Camerer. 1994. The Predictive Utility of Generalized Expected Utility
Theories. Econometrica 62 1251-1289.

Hart, O., D. Jaffee. 1973. On the Application of Portfolio Theory to Depository Financial
Intermediaries. Review of Economic Studies 41 129-147.

Hawawani, G. 1978. A Mean-Standard Deviation Exposition of the Theory of the Firm

under Uncertainty: A Pedagogical Note. American Economic Review 68 194-202.

Heath, C., A. Tversky. 1991. Preference and Belief: Ambiguity and Competence in
Choice under Uncertainty. Journal of Risk and Uncertainty 4 5-28.

Heath, C., S. Huddart, M. Lang. 1999. Psychological Factors and Stock Option Exercise.
The Quarterly Journal of Economics 114 601-628.

Hershey, J., P. Schoemaker 1980. Risk Taking and Problem Context in the Domain of
Losses: An Expected Utility Analysis. Journal of Risk and Insurance 47 111-132.

Hershey, J., P. Schoemaker. 1985. Probability vs Certainty Equivalence Methods in

Utility Measurement: Are they Equivalent. Management Science 31 1213-1231.

Herstein, I., J. Milnor 1953. An Axiomatic Approach to Measurable Utility.

Econometrica 21 291-297.

Hey, J. 1997. The Economics of Uncertainty I. Edward Elgar Publishing Limited, U.K.

Hey, J. 1997. The Economics of Uncertainty II. Edward Elgar Publishing Limited, U.K.

Hey, J.D., C. Orme 1994. Investigating generalizations of Expected Utility Theory Using
Experimental Data. Econometrica 62 1291-1326.

Hicks, J. 1979. Causality in Economics. Basil Blackwell, Oxford.

Hodgkinson G., N. Bown, A. Maule, K. Glaister, A. Pearman 1999. Breaking the frame:
An Analysis of Strategic Cognition and Decision Making under Uncertainty.
Strategic Management Journal 20 977-985.

Huxley, A. 1945. The Perennial Philosophy. Harper and Row Publishers, New York.

229
Ignatieff, M. 1998. A Life: Isaiah Berlin. Henry Holt and Company, New York.

Jensen, N. 1967. An Introduction to Bernoullian Utility Theory. I. Utility Functions.

Swedish Journal of Economics 69 163-183.

Jevons, W.S. 1871. Theory of Political Economy. Edition 1957. Kelley and MacMillan,
NewYork.

Jullien, B., B. Salanie. 2000. Estimating Preferences under Risk: The Case of Racetrack
Bettors. Journal of Political Economy 108 503-530.

Kahneman, D., A. Tversky 1979. Prospect Theory: An Analysis of Decision under Risk.
Econometrica 47 263-291.

Kahneman, D., A. Tversky. 2000. Choices, Values and Frames. Cambridge University
Press, Cambridge.

Kahneman, D., J. Knetsch, R. Thaler 1991. Anomalies: The Endowment Effect, Loss
Aversion and Status Quo Bias. D. Kahneman and A. Tversky, eds. Choices,
Values and Frames. Cambridge University Press, Cambridge.

Kamarkar, U. 1979. Subjectively Weighted Utility: A Descriptive Extension of the

Expected Utility Model. Organizational Behavior and Human Performance 47
263-293.

Karni, E., D. Schmeidler. 1991. Utility Theory with Uncertainty. W. Hildenbrand and H.
Sonnenschein, eds. Handbook of Mathematical Economics. North-Holland,
Amsterdam.

Kleinfield, S. 1983. The Traders. Holt, Rinehart and Winston, New York.

Knetsch, J. 1990. The Endowment Effect and Evidence of Nonreversible Indifference

Curves. D. Kahneman and A. Tversky. eds. Choices, Values and Frames.
Cambridge University Press, Cambridge.

Köbberling, V., P.P. Wakker. 2005. An Index of Loss Aversion. Journal of Economic
Theory 122 119-131.

Kreps, D. 1988. Notes on the Theory of Choice. Westview Press, U.S.

230
Kuhn, T. 1970. The Structure of Scientific Revolutions. University of Chicago Press
Chicago.

Lattimore P., J. Baker, A. Witte. 1992. The Influence of Probability on Risky Choice: A
Parametric Estimation. Journal of Economic Behavior and Organization 17 377-
400.

Levy, H. 1992. Stochastic Dominance and Expected Utility: Survey and Analysis.
Management Science 38 555-593.

Levy, M., H. Levy. 2002. Prospect Theory: Much Ado about Nothing. Management
Science 48 1334-1349.

Levy, H., Z. Wiener. 1998. Stochastic Dominance and Prospect Dominance with
Subjective Weighting Functions. Journal of Risk and Uncertainty 16 147-163.

Lichtsenstein, S., P. Slovic. 1971. Reversals of Preferences between Bids and Choices in
Gambling Decisions. Journal of Experimental Psychology 89 46-55.

Lindman, H. 1971. Inconsistent Preferences among Gambles. Journal of Experimental

Psychology 80 590-597.

Loewenstein, G., D. Adler. 1995. A Bias in the Prediction of Tastes. The Economic
Journal 105 929-937.

Loewenstein, G. 1988. Frames of Mind in Intertemporal Choice. Management Science 34

200-214.

Loomes, G., R. Sugden. 1983. Regret Theory: An Alternative Theory of Rational Choice
under Uncertainty. The Economic Journal 92 805-824.

Lopes, L. 1987. Between Hope and Fear: The Psychology of Risk. Advances in
Experimental Psychology 20 255-295.

Lopes, L., G. Oden. 1999. The Role of Aspiration Level in Risky Choice: A Comparison
of Cumulative Prospect Theory and SP/A Theory. Journal of Mathematical
Psychology 43 286-313.

Luce, D. 2000. Utility of Gains and Losses: Measurement-Theoretic and Experimental

Approaches. Lawrence Erlbaum Associates, Inc. Mahwah, New Jersey.

231
Luce, D., H. Raiffa. 1957. Games and Decisions. Wiley and Sons, NewYork.

Luce, D., P. Fishburn. 1991. Rank and Sign-Dependent Linear Utility models for Finite
first-Order Gambles. Journal of Risk and Uncertainty 4 25-29.

Luce, D. 1991. Rank and Sign-Dependent Linear Utility models for Binary Gambles.
Journal of Economic Behavior 53 75-100.

MacCrimmon, K., S. Larsson. 1979 Utility Theory: Axioms versus Paradoxes. M. Allais
and O. Hagen, eds. Expected Utility Hypothesis and the Allais Paradox. Reidel,
Dordrecht/Boston.

Machina, M. 1989. Dynamic Consistency of Non-Expected Utility Models of Choice

under Uncertainty. Journal of Economic Literature 27 1662-1688.

Machina, M., D. Schmeidler. 1992. A More Robust Definition of Subjective Probability.

Econometrica 60 745-780.

Machina, M. 1982. Expected Utility Analysis without the Independence Axiom.

Econometrica 50 277-323.

Markowitz, H. 1952. The Utility of Wealth. Journal of Political Economy 60 151-158.

May, K. 1954. Intransitive Utility and Aggregation of Preference Patterns. Econometrica

50 277-323.

McClennen, E. 1990. Rationality and Dynamic Choice. Cambridge University Press,

Cambridge.

McCord, M., R. de Neufville. 1986. Lottery Equivalents: Reduction of the Certainty

Effect Problem in Utility Assessment. Management Science 32 56-60.

Mehra, R., E. Prescott. 1985. The Equity Premium: A Puzzle. Journal of Monetary
Economics 15 145-161.

Menger, C. 1871. Grundsätze der Volkwirtschaftslehre. Translated into English, 1971.

Principles of Economics. Ed. 1981, New York University Press, New York.

Montaigne, M. 1588. Essais, Book I, Vol.18, de La Peur.

232
Munier, B. 1989. Calcul Economique et Révision de la Théorie de la Décision en Avenir
Risqué. Revue d’Economie Politique 2 276-306.

Nau, R. 2004. Ph.D. Seminar on Choice Theory. The Fuqua Business School. Webpage:
www. [Link].

Neilson, W. 2002. Comparative Risk Sensitivity with Reference-Dependent Preferences.

Journal of Risk and Uncertainty 24 131-142.

Neilson, W. 1992. A Mixed Fan Hypothesis and its Implications for Behavior toward
Risk. Journal of Economic Behavior Organization 19 197-211.

Odean, T. 1998. Are Investors Reluctant to Realize their Losses? The Journal of Finance
53 1775-1798.

Officer, R., A. Halter. 1968. Utility Analysis in a Practical Setting. American Journal of
Agricultural Economics 50 257-277.

Pareto, V. 1906. Manuel d’Economie Politique. Translated into English, 1927. Manual of
Political Economy. Ed 1971, MacMillan Press, London.

Pascal, B. 1963. Oeuvres Complètes. Présentation et Notes de Lafuma, L. Paris, Seuil.

Payne, J. 2005. It Is whether You Win or Lose: The Importance of Overall Probabilities
of Winning or Losing in a Risky Choice. Journal of Risk and Uncertainty 30 5-
19.

Payne, J., Bettman, J., D. Schkade. 1999. Measuring Constructed Preferences: Towards a
Building Code. Journal of Risk and Uncertainty 19 243-270.

Payne, J.W., D.J. Laughhunn, R. Crum. 1981. Further Tests of Aspiration level Effects in
Risky Choice. Management Science 27 953-958.

Pennings, J., A. Smidts. 2003. The Shape of Utility Functions and Organizational
Behavior. Management Science 49 1251-1263.

Platt, M., P. Glimcher. 1999. Neural Correlates of Decision Variables in Parietal Cortex.
Nature 400 233-238.

Popkin, S. 1992. The Reasoning Voter. The University of Chicago Press, Chicago.

233
Pratt, J. 1964. Risk Aversion in the Small and in the Large. Econometrica 32 122-136.

Prelec, D. 1990. A ‘Pseudo-Endowment’ Effect and its Implications for Some Recent
Non-Expected Utility Models. Journal of Risk and Uncertainty 3 247-259.

Prelec, D. 1998. The Probability Weighting Function. Econometrica 66 497-527.

Prelec, D. 2000. Compound Invariant Weighting Functions in Prospect Theory. D.

Kahneman and A. Tversky, eds. Choices, Values and Frames. Cambridge
University Press, Cambridge.

Preston M., P. Baratta. 1948. An Experimental Study of the Auction-Value of an

Uncertain Outcome. American Journal of Psychology 61 183-193.

Putler, D. 1992. Incorporating Reference Price Effects into a Theory of Consumer

Choice. Marketing Science 11 287-309.

Quiggin, J. 1982. A Theory of Anticipated Utility. Journal of Economic Behavior and

Organization 3 323-343.

Quiggin, J. 1991. Comparative Statics for Rank-dependent Expected Utility Theory.

Journal of Risk and Uncertainty 4 339-350.

Quiggin, J. 1991. On the Optimal Design of Lotteries. Economica 58 1-16.

Quiggin, J. 2004. www. [Link]. economics/johnquiggin.

Rabin M., R. Thaler. 2001. Anomalies: Risk Aversion. Journal of Economic Perspectives
15 219-232.

Rabin, M. 2000. Risk Aversion and Expected-Utility Theory: A Calibration Theorem.

Econometrica 68 1281-1292.

Ramsey, F. 1931. Foundations: Essays in Philosophy, Logic, Mathematics and

Economics. Routledge and Kegan Paul, London.

Rothschild, M., J. Stiglitz. 1970. Increasing Risk: I. A Definition. Journal of Economic

Theory 2 225-243.

234
Rothschild, M., J. Stiglitz. 1971. Increasing Risk: II. Its Economic Consequences.
Journal of Economic Theory 3 66-84.

Rottenstreich, Y., A. Tversky. 1997. Unpacking, Repacking, and Anchoring: Advances in

Support Theory. Psychological Review 104 406-415.

Rustichini, A. et al. Forthcoming. Games and Economic Behavior.

Saha, A. 1993. Expo-Power Utility: A ‘Flexible’ form for Absolute and Relative Risk
Aversion. American Journal of Agricultural Economics 75 905-913.

Samuelson, P. 1963. Risk and Uncertainty: A Fallacy of Large Numbers. Scientia 98

108-113.

Samuelson, W., R. Zeckhauser. 1988. Status Quo Bias in Decision Making. Journal of
Risk and Uncertainty 1 7-59.

Savage, L.J. 1954. The Foundations of Statistics. John Wiley and Sons, New York.

Schkade, D. 1988. Toward Respondent-Centered Environmental Valuation: Asking

Questions that People Can Answer. Working Paper, Department of Management,
University of Texas, Austin, TX.

Schmeidler, D. 1989. Subjective Probability and Expected Utility without Additivity.

Econometrica 57 571-587.

Schmidt, U. 2001. Lottery-Dependent Utility: A Re-Examination. Theory and Decision

50 35-58.

Schmidt, U. 2002. Alternatives to Expected Utility: Some Formal Theories. P.J.

Hammond, S. Barberá and C. Seidl, eds. Handbook of Utility Theory, II,
forthcoming, Boston: Kluwer, Boston.

Schmidt, U. 2003. Reference Dependence in Cumulative Prospect Theory. Journal of

Mathematical Psychology 47 122-131.

Schmidt, U. 2003. The Axiomatic Basis of Risk-Value models. European Journal of

Operational Research 145 216-220.

235
Schmidt, U., S. Traub. 2002. An Experimental Test of Loss Aversion. Journal of Risk
and Uncertainty 25 233-249.

Schmidt, U., H. Zank. 2001. A New Examination of Rank-Dependent Utility with Trade-
off Consistency for Equally Likely Events. Journal of Mathematical Economics
35 483-491.

Schmidt, U. and H. Zank. 2002. Strong Risk Aversion in Cumulative Prospect Theory.
Working Paper, The University of Manchester, U.K.

Schmidt, U., H. Zank. 2005. What is Loss Aversion. Journal of Risk and Uncertainty 30
157-167.

Schneider, S. L., L. Lopes. 1986. Reflections in Preferences under Risk: Who and When
May Suggest Why. Journal of Experimental Psychology: Human Perception and
Performance 12 535-548.

Schoemaker, P., H. Kunreuther. 1979. An Experimental Study of Insurance Decisions.

Journal of Risk and Insurance 46 603-618.

Segal, U. 1989. Anticipated Utility: A Measure Representation Approach. Annals of

Operation Research 19 359-374.

Segal, U. 1990. Two-Stage Lotteries without the Reduction Axiom. Econometrica 58

349-377.

Shackle, G. 1955. Uncertainty in Economics and Other Reflections. Cambridge

University Press, Cambridge:).

Shackle, G. 1969. Decision Order and Time in Human Affairs. Cambridge University
Press Cambridge.

Shalev, J. 2000. Loss Aversion Equilibrium. International Journal of Game Theory 29

269-287.

Shefrin, H., M. Statman. 1985. The Disposition to Sell Winners too Early and to Ride
Losers too Long: Theory and Evidence. Journal of Finance 40 777-790.

Simon, H. 1954. Review of vNM’s Book: Theory of Games and Economic Behavior.
American Journal of Sociology 50 559-560.

236
Sinn, H.-W. 1983. Economic Decision Under Uncertainty. North Holland Publishing
Company, Holland).

Slovic, P. 1969. Manipulating the Attractiveness of a Gamble without Changing its

Expected Value. Journal of Experimental Psychology 79 139-145.

Slovic, P., S. Lichtenstein. 1983. Preference Reversals: A Broader Perspective. American

Economic Review 73 596-605.

Smith, A. 1776. An Inquiry into the Nature and Causes of the Wealth of Nations. E.
Cannan, ed. Edition 1904 Methuen and Co, London.

Smith, K. et al. 2002. Neuronal Substrates for Choice under Ambiguity, Risk, Certainty,
Gains and Losses. Management Science 48 182-189.

Spinoza, B. 1677. Ethics. Translated into English, E. Curley, 1989. The Collected Works
of Spinoza. Princeton University Press, Princeton, New Jersey, U.S.

Starmer, C. 1992. Testing New Theories of Choice under Uncertainty Using the Common
Consequence Effect. Review of Economic Studies 59 813-830.

Starmer, C. 1999. Experiments in Economics: Should we Trust the Dismal Scientists in

White coats? Journal of Economic Methodology 6 1-30.

Starmer, C. 2000. Developments in Non-Expected Utility Theory: The Hunt for a

Descriptive Theory of Choice under Risk. Journal of Economic Literature 38
332-382.

Stevens, S.S. 1970. Neural Events and the Psychophysical Law. Science 170 1043-1050.

Stout, G.F. 1931. Mind and Matter. London.

Sugden, R. 2003. Reference-Dependent Subjective Expected Utility. Journal of

Economic Theory 111 172-191.

Swalm, R.O. 1966. Utility Theory-Insights into Risk Taking. Harvard Business Review
44 123-136.

Thaler, R. 1980. Toward a Positive Theory of Consumer Choice Journal of Economic

Behavior and Organization 1 39-60.

237
Thaler, R., E.J. Johnson. 1990. Gambling with the House Money and Trying to Break
Even: The Effects of Prior outcomes on Risky Choice. Management Science 36
643-660.

Thaler, R. 1993. Advances in Behavioral Finance. Russell Sage Foundation, New York.

Thaler, R. et al. 1997. The Effect of Myopia and Loss Aversion on Risk Taking: An
Experimental Test Quarterly Journal of Economics 112 647-661.

Tobin, J. 1958. Liquidity Preference as Behavior towards Risk. The Review of Economic
Studies 25 65-86.

Tversky, A. 1969. Intransitivity of Preferences. Psychological Review 76 31-48.

Tversky, A., D. Kahneman. 1992. Advances in Prospect Theory: Cumulative

Representation of Uncertainty. Journal of Risk and Uncertainty 5 297-323.

Tversky, A., C. Fox. 1995. Weighting Risk and Uncertainty. Psychological Review 102
269-283.

Tversky, A., D. Kahneman. 1981. The Framing of Decisions and the Psychology of
Choice. Science 211 453- 458.

Tversky, A., D. Kahneman. 1986. Rational Choice and the Framing of Decisions. D.
Kahneman and A. Tversky, eds. Choices, Values and Frames. Cambridge
University Press, Cambridge.

Tversky, A., D. Kahneman. 1991. Loss Aversion in a Riskless choice: A Reference

Dependent model. D. Kahneman and A. Tversky, eds. Choices, Values and
Frames. Cambridge University Press, Cambridge.

Tversky, A., D. J. Koehler. 1994. Support theory: A Non-Extentional Representation of

Subjective Probability. Psychological Review 101 547–567.

Tversky, A., P.P. Wakker. 1995. Risk Attitudes and Decision Weights. Econometrica 63
1255-1280

Tversky, A., P. Slovic, P., S. Sattath. 1988. Contingent Weighting in Judgment and
Choice. Psychological Review 95 371-384.

238
Tversky, A., P. Slovic, D. Kahneman. 1990. The Causes of Preference Reversals.
American Economic Review 80 204-217.

Vind, K. 2003. Independence, Additivity, Uncertainty. Springer Verlag, Berlin.

von Neumann, J. and O. Morgenstern. 1944. Theory of Games and Economic Behavior
Princeton. Princeton University Press, Princeton.

von Winterfeldt, D., W. Edwards. 1986. Decision Analysis and Behavioral Research.
Cambridge University Press, Cambridge.

Wakker, P. P. 1994. Separating Marginal Utility and Probabilistic Risk Aversion. Theory
and Decision 36 1-44.

Wakker, P. P. 1999. Justifying Bayesianism by Dynamic Decision Principles and Belief.

Working Paper, Leiden University, The Netherlands.

Wakker, P. P. 2001. Testing and Characterizing Properties of Non-Additive Measures

through Violations of the Sure-Thing Principle. Econometrica 69 1039-1059.

Wakker, P. P. 2004. On the Composition of Risk Preference and Belief. Psychological

Review 111 236-241.

Wakker, P. P. 2003. The Data of Levy and Levy (2002) Actually Support Prospect
Theory. Management Science 49 979-981.

Wakker, P. P., A. Tversky. 1993. An Axiomatization of Cumulative Prospect Theory.

Journal of Risk and Uncertainty 7 147-176.

Wakker, P., R. Thaler, A. Tversky, A. 1997. Probabilistic Insurance. Journal of Risk and
Uncertainty 15 7-28.

Wakker, P.P., D. Deneffe. 1996. Eliciting von Neumann-Morgenstern Utilities when

Probabilities are Distorted or Unknown. Management Science 42 1131-1150.

Walras, L. 1874. Elements d’Economie Pure. ed. 1926. Translated into English, 1954.
Elements of Pure Economics or the Theory of Social Wealth. Richard Irwin,
Homewood, Illinois, U.S.

239
Weber, E. 1994. From Subjective Probabilities to Decision Weights: The Effects of
Asymmetric Loss Functions on the Evaluation of Uncertain Outcomes and
Events. Psychological Bulletin 115 228-242.

Weber, E., W. P. Bottom. 1989. Axiomatic Measures of Perceived Risk: Some Tests and
Extensions. Journal of Behavioral Decision Making 2 113-131.

Weber, M., C. Camerer. 1988. The Disposition Effect in Securities Trading: An

Experimental Analysis. Journal of Economic Behavior and Organization 33 167-
184.

Wilber, K. 1996. A Brief History of Everything. Shambhala Publications, U.S.

Wollheim, R. 1999. On the Emotions. Yale University Press, London.

Wu, G. and R. Gonzalez. 1996. Curvature of the Probability Function. Management

Science 42 1676-1690.

Wu, G., Gonzalez, R. 1998. Common Consequence Conditions in Decision Making

under Risk. Journal of Risk and Uncertainty 16 115-139.

Wu, G., R. Gonzalez. 1999. Nonlinear Decision Weights in Choice under Uncertainty.
Management Science 45 74-85.

Yaari, M.E. 1987. The Dual Theory of Choice under Risk. Econometrica 55 95-115.

240
APPENDIX A: DATA FIGURES

Figure 8 : The Certainty Equivalent for a Concave Utility Function

241
APPENDIX B: DISPLAY

Display II: Illustration of the Use of the Scrollbar

242
APPENDIX C: DATA TABLES

Table 5: Classification of Alternative Theories

Models Indifference Curves Probabilities

vNM (1944) Parallel, and Linear Objective
Machina (1982) Smooth, not necessarily Linear, and Fan-out Objective
Chew and MacCrimmon (1979) Linear, not Parallel, and Fan-out Objective
Gul-Neilson (1992) Linear, Fan-in, and Fan-out Objective
Chew,Segal and Epstein (1991) Not Linear, Mixed Fanning, May switch from Objective
concave to convex
Quiggin (1982) Concave, Fan-out, and Parallel at Hypotenuse Decision Weights
Kahneman and Tversky (1992) Not-Linear and Mixed Fanning Decision Weights

Table 8: The Practitioners’ Institutions and their Locations

Financial Institution Location
Martingale Asset Manag. Boston- US
GMT Capital Corp. Atlanta-US
GMT Capital Corp. Atlanta-US
Boyd Watterson Cleveland-US
Boyd Watterson Cleveland-US
Lorain National Bank Cleveland-US
Lorain National Bank Cleveland-US
Smith Barney Cleveland-US
Smith Barney Cleveland-US
Smith Barney Cleveland-US
Smith Barney Cleveland-US
Smith Barney Cleveland -US
UBS PaineWebber Cleveland-US
Butler Wick &Co., Inc Cleveland-US
Smith Barney New York NewYork-US
Smith Barney New York NewYork-US
Smith Barney New York NewYork-US
Smith Barney New York NewYork-US
Corey Capital Inc. Phoenix-US
Corey Capital Inc. Phoenix- US
Corey Capital Inc. Phoenix- US

243
 Wells Fargo Phoenix-US
Wells Fargo Phoenix- US
Coble Pension &Wealth
Ma. Phoenix-US
First National Bank Beirut-Lebanon
Saradar Bank Beirut-Lebanon
Allied Bank Beirut-Lebanon
Byblos Bank Beirut-Lebanon
Comgest Beirut-Lebanon
Financial Funds Advisors Beirut-Lebanon
Audi Bank Beirut-Lebanon
Capital Trust Beirut-Lebanon
Middle East Capital Group Beirut-Lebanon
Audi Bank Beirut-Lebanon
Societe Generale - Fidus Beirut-Lebanon
Capital Trust Beirut-Lebanon
Intercontinental Bank Beirut-Lebanon
Meryll Lynch Beirut-Lebanon
Credit Agricole Indo- Suez Beirut-Lebanon
Meryll Lynch Beirut-Lebanon
Credit Agricole Indo- Suez Beirut-Lebanon
Audi Bank Beirut-Lebanon
Saoudi Lebanese Bank Beirut-Lebanon
El-Rashed Investment Beirut-Lebanon
Lebanese Canadian Bank Beirut-Lebanon
Banque du Liban et D'outre
mer Beirut-Lebanon

Table 10: Exponential and Power Risk Aversion Coefficients

Risk Risk Relative Risk

Utility Aversion Tolerance Aversion
Function u(x) -u''(x)/u'(x) -u'(x)/u''(x) -x u''(x)/u'(x)
Exponential u(x) = -e-rx r (CARA) 1/r rx
c
Power u(x) =1/c (x) (1-c)/(x) (x)/(1-c) 1-c (CRRA)
c < 1, c ≠ 0

244
 Table 16: Main Parameters for Practitioners

Practitioner Probabilities Power Exponential Expo-Power

Number P1 P2 Alpha Beta Alpha Beta Alpha Beta
2 61 41 2.380998 0.441945 -2.51716 3.554083 2.7409045 0.663343
3 23 87 0.432224 1.882478 2.265403 -1.8518 0.6711281 2.152872
4 71 51 1.215347 0.882535 -0.68419 0.753242 1.4867638 1.298799
5 63 61 1.105859 0.559464 -0.27301 2.885824 1.3764401 0.764835
6 93 35 1.486456 0.624727 -1.25701 1.412348 1.795828 0.896381
7 88 71 1.011 0.722583 0.048198 1.351406 1.2730101 0.949645
8 58 51 1.471043 0.930471 -1.19596 0.242682 1.7700915 1.214302
9 83 37 0.874757 0.773768 0.451493 1.026298 1.1323114 1.000048
10 54 57 0.809379 1.084596 0.583785 -0.27368 1.0706879 1.331851
11 28 57 0.670522 8.715305 1.610599 0.901659
12 24 6 0.550002 0.382677 2.343562 3.95954 0.7857909 0.617033
13 87 41 0.696171 0.940786 1.175854 0.150988 0.9441502 1.236092
14 56 55 1.189312 0.58929 -0.52336 1.818705 1.4732133 0.827597
15 24 70 0.640664 1.447426 0.928459 -1.1093 0.9214748 1.742205
16 43 52 0.730674 1.709349 1.065489 -1.65546 0.9792663 1.99568
17 71 39 1.054633 0.682597 -0.14255 1.348408 1.3241301 0.9244
18 75 14 0.539749 0.705178 2.592722 1.213546 0.7724195 0.961374
19 86 8 1.032571 1.086253 -5.52004 -0.16608 1.2385047 1.363351
20 37 55 0.757344 0.680393 0.970483 1.432612 1.0105909 0.913767
21 29 35 0.504055 0.943881 2.805818 0.223523 0.7352558 1.204265
22 36 77 0.531561 1.827176 1.925866 -1.89741 0.7851091 2.149494
23 66 34 0.58038 0.317038 2.100867 5.43035 0.8193522 0.541587
24 22 84 0.43282 1.717576 3.691506 -1.47282 0.6565923 2.02411
25 25 78 0.368992 6.061721 4.357306 -6.39013 0.5849558 6.451708
26 65 85 0.762166 1.675184 0.894451 -1.54494 1.0180301 1.959153
27 32 34 0.531043 1.114002 2.315585 -0.33631 0.7664623 1.401521
28 26 87 0.841566 0.37596 0.421047 3.663079 1.1246377 0.622335
29 21 48 1.48548 9.150792 -1.16162 1.757882
30 69 63 1.63816 1.082252 -1.49885 -0.20868 1.9244205 1.362707
31 19 34 0.553635 0.529052 2.310261 2.158697 0.7885674 0.77949
32 31 56 0.61065 0.80902 1.867491 0.702876 0.8464555 1.073669
33 71 45 1.876947 0.759428 -1.8342 0.967491 2.1957023 1.01301
34 40 30 0.843514 0.294514 0.509515 5.44487 1.109064 0.517718
35 80 70 0.497438 1.677583 3.03907 -1.62348 0.72059 1.948343
36 89 96 0.790695 2.118735 0.899013 -2.13339 1.0341132 2.443526
37 51 42 0.912529 0.903647 0.309114 0.347053 1.1774802 1.164783
38 98 20 0.425302 5.383797 4.393571 0.639634
39 68 31 0.57544 1.556574 1.99136 -1.27005 0.8175628 1.865881
40 79 32 0.58038 1.725289 2.100867 -1.54789 0.8193522 2.023833
41 78 34 0.996992 -0.086 5.657763 1.254488 0.562262
42 56 44 1.042085 0.578304 -0.21605 1.902523 1.316069 0.820593

245
43 79 66 0.948722 0.7453 0.090909 1.231125 1.229641 0.968369
44 80 35 0.608574 2.119463 1.697247 -2.12598 0.8566023 2.44121
45 86 57 0.579813 0.848784 2.07892 0.499637 0.814938 1.110919
46 73 28 0.577194 0.900487 2.145681 0.323158 0.8145957 1.171631
47 77 42 0.622703 1.010525 1.768889 0.8596306 1.281592

246
APPENDIX D: The Questionnaire

The following hypothetical choices are designed to investigate your attitude towards risk.
Try to be as accurate as possible in predicting your choices. The responses are
anonymous and there is no correct answer, hence, no reason not to state your true
preference. After each choice you make, please state why you chose it in the blank box
below.

A- Suppose that you decided to invest $ 10000 either in stock F or in stock G. Which
stock would you choose, F, or G when it’s given that the dollar gain or loss one month
from now will be as follows.

Task I:
F G
Gain/Loss Probability Gain/Loss Probability
-6000 1/4 -3000 1/2
3000 3/4 4500 1/2

Please write F or G :
Please state the reason for your choice:

Task II:
F G
Gain/Loss Probability Gain/Loss Probability
-1600 1/4 -1000 1/4
-200 1/4 -800 1/4
1200 1/4 800 1/4
1600 1/4 2000 1/4

Please write F or G :
Please state the reason for your choice:

Task III:
F G
Gain/Loss Probability Gain/Loss Probability
-3000 1/4 -1500 1/2
3000 3/4 4500 1/2

Please write F or G :
Please state the reason for your choice:

247
Task IV:
F G
Gain/Loss Probability Gain/Loss Probability
-5000 1/6 -5000 1/6
-3000 1/6 -2000 1/6
-500 1/6 -1500 1/6
2000 1/6 1000 1/6
3000 1/6 4000 1/6
5000 1/6 5000 1/6
Please write F or G :
Please state the reason for your choice:

Task V:
F G
Gain/Loss Probability Gain/Loss Probability
-3000 30% -3000 10%
3000 60% -1500 40%
4500 10% 4500 50%

Please write F or G :
Please state the reason for your choice:

Task VI:
F G
Gain/Loss Probability Gain/Loss Probability
-6000 10% -6000 10%
-3000 20% -1500 40%
3000 60% 4500 40%
6000 10% 6000 10%

Please write F or G :
Please state the reason for your choice:

248
Task VII:
F G
Gain/Loss Probability Gain/Loss Probability
-6000 1/3 -6000 1/6
3000 1/2 -3000 1/3
4500 1/6 4500 1/2

Please write F or G :
Please state the reason for your choice:

Task VIII:
F G
Gain/Loss Probability Gain/Loss Probability
-6000 30% -6000 10%
3000 60% -3000 40%
4500 10% 4500 50%

Please write F or G :
Please state the reason for your choice:

Task IX:
F G
Gain/Loss Probability Gain/Loss Probability
-6000 26% -6000 2%
3000 72% -3000 48%
4500 2% 4500 50%

Please write F or G :
Please state the reason for your choice:

Task X:
F G
Gain/Loss Probability Gain/Loss Probability
1000 1/2 0 1/2
2000 1/2 3000 1/2

Please write F or G :
Please state the reason for your choice:

249
Task XI:
F G
Gain/Loss Probability Gain/Loss Probability
0 10% 0 50%
1000 40%
2000 40%
3000 10% 3000 50%

Please write F or G :
Please state the reason for your choice:

Task XII:
F G
Gain/Loss Probability Gain/Loss Probability
-3000 1/2 -2000 1/2
0 1/2 -1000 1/2

Please write F or G :
Please state the reason for your choice:

Task XIII:
F G
Gain/Loss Probability Gain/Loss Probability
-3000 50% -3000 10%
-2000 40%
-1000 40%
0 50% 0 10%

Please write F or G :
Please state the reason for your choice:

Task XIV:
F G
Gain/Loss Probability Gain/Loss Probability
-1000 10% -1000 50%
0 80%
1000 10% 1000 50%

Please write F or G :
Please state the reason for your choice:

250
Task XV:
F G
Gain/Loss Probability Gain/Loss Probability
-3000 20% -3000 50%
0 60%
3000 20% 3000 50%

Please write F or G :
Please state the reason for your choice:

Task XVI:
F G
Gain/Loss Probability Gain/Loss Probability
-3000 20% -3000 50%
-1000 30%
1000 30%
3000 20% 3000 50%

Please write F or G :
Please state the reason for your choice:

Task XVII:
F G
Gain/Loss Probability Gain/Loss Probability
-500 10% -500 30%
0 40% 500 20%
1500 40% 1000 20%
2000 10% 2000 30%

Please write F or G :
Please state the reason for your choice:

251
Task XVIII:
F G
Gain/Loss Probability Gain/Loss Probability
-2000 30% -2000 10%
-1000 60%
0 60%
1000 10% 1000 30%

Please write F or G :
Please state the reason for your choice:

Task XIX:
F G
Gain/Loss Probability Gain/Loss Probability
-5000 15% -5000 35%
-3000 30% -1000 30%
0 20%
3000 20%
5000 15% 5000 35%
Please write F or G :
Please state the reason for your choice:

Task XX:
F G
Gain/Loss Probability Gain/Loss Probability
-1500 20% -1500 50%
1500 60%
4500 20% 4500 50%

Please write F or G :
Please state the reason for your choice:

B-Consider now an investment whose possible outcomes and their probabilities are the
following:

Gain/Loss Probability
100 0.2
50 0.2
0 0.2
-25 0.2
-50 0.2
1) If you could add a sum of money ($38) to either the outcome that paid $100 or the
outcome that paid $ 0, which outcome would you choose?

252
Please write your answer:
Please state the reason for your choice:

2) If you could add a sum of money ($38) to either the outcome that paid $100 or the
outcome that paid $ 50, which outcome would you choose?
Please write your answer:
Please state the reason for your choice:

253
LIST OF FIGURES

Figure 1 The Independence Axiom 51

Figure 2 Indifference Curves Assuming EU 61

Figure 3 Indifference Curves Assuming PT 64

Figure 4 Indifference Curves Assuming RDU 67

Figure 5 Indifference Curves Assuming CPT 70

Figure 6 The Utility Function Assuming CPT 77

Figure 7 The Weighting Function Assuming CPT 79

Figure 8 The Certainty Equivalent for a Concave Utility Function 81

Figure 9 Illustration of the Elicitation Procedure 182

Figure 10 The Shape of the Utility for the Median Practitioner 121

Figure 11 Prospect and Markowitz Utility functions 133

Figure 12 The Cumulative Probability Distributions for Task I 136

Figure 13 The Wcd Class Probability Weighting Function 139

Figure 14 The Cumulative Probability Distributions for Task VII 140

Figure 15 Stochastic Dominance Accounting for Task XIX 142

LIST OF TABLES

Table 1 The Allais Common Consequence Sets 45

Table 2 The Allais Common Ratio Sets 47

Table 3 The Allais Paradox in Decumulative Probabilities 48

Table 4 The Ellsberg Paradox 53

254
Table 5 Classification Table of Alternative Theories 94

Table 6 Characterization of Risk Aversion under the Different Theories 105

Table 7 Four-Steps Elicitation Procedure 185

Table 8 Practitioners’ Institutions and their Locations 186

Table 9 An Illustration of the Bisection Method 187

Table 10 Exponential and Power Risk Aversion Coefficients 120

Table 11 Non-Parametric Classification of the Practitioners 121

Table 12 Parametric Fittings for the Utility Function for Gains and Losses 122

Table13 (a) Parametric Classification in Number and Proportion 123

Parametric Crossing of Gains by Losses for the Power,

the Exponential and the Expo-Powers

Table 14 Individual Loss Aversion Results 124

Table 15 Comparison of the Mean (Median) of Practitioners/Students 124

Table 16(b) Other Parameters for Practitioners 188

Table 17 Results of Experiment II Regarding the Shape of U 147

Table 18 Results for Tasks XIV-XVI 148

Table 19 Results for Tasks XVII-XX 149

Table 20 Results and P-Values for Binomial for Experiment II 191

255
256