UNDERSTANDING MARKETS AND

CUSTOMERS
Outline of the presentation

• Introductory hypothesis
• Identifying the market
• Customer choice
• Market research
• Product positioning
 Introductory hypothesis
• We suppose to be operating in a market-pull product
development process (customer-driven product
development)
• Mainstream approach since the ‘90s… but being constantly
updated, given today’s environment, characterized by
• Radically innovative / paradigm changing products → can use the
same tools, but in a different way
• Highly complex products, product-services and platform-based
services based on agile/lean innovation → must also use big data /
analytics
• Design-driven innovation (e.g. Artemide, B&O, etc.) → must also
manage the interaction between firms' proposals and customers'
tacit needs (functional and/or emotional)
Introductory hypothesis
• Understanding the market is at the core of product
development, especially for consumer products (B2C)
• Producers have few contacts with final customers → need to
have a deep knowledge of
• needs and wants
• purchasing processes (simple but tacit)
• Products are chosen by taking many criteria into account (not
only technical performance)
• The product has to be developed as a complex system
(product, marketing strategy, complementary products)
• The brand defines a coherent set of features and
communicates it
• Within the firm
• Towards the market
Introductory hypothesis

• For B2B firms

• Market research is less important
• Fewer number of customers
• Customers are “closer” to the firm
• Purchasing behavior is dominated by technical criteria
• The purchasing process is more complex (but "visible")
• With very innovative products (technology push
prevails) market research
• is less important
• covers different role in the product development process
Identifying the market

• It is the first step of marketing activities (non-trivial)

• Identification implies defining what is in and what is out
• There is a number of possible criteria
Identifying the market

• Criteria from economics – a set of products belongs to

the same market…
• If they are substitutes (positive cross-price demand elasticity)
Qy Qy
0
Px Px
• Prices are strongly correlated (which implies competitive
dynamics)
• If producers could potentially act as a cartel and bring the
prices to monopoly level
Identifying the market

• Criteria from economics are difficult to apply

• Reliability of econometric estimates
• Defining thresholds
• “what-if” hypothesis are difficult to verify
• In practice
• decisions depend on the specific objectives of the research
• firms tend to make hierarchical decisions (a broader “market”,
and narrower “segments”)
Identifying the market

• Criteria based on hierarchical industry classification

systems
• ISTAT (ATECO07)
• Eurostat (NACE)
• USA (NAICS )
• Geography (if arbitrage is made difficult by
transportation cost, national norms, customs, etc.)
• Similarity in technical features
• Similarity in usage patterns and needs being satisfied
(outcome-based segmentation)
Identifying the market

• Explicit evaluations coming from a sample of customers

(e.g. brand switching behavior)
• Demographic criteria
• Membership to predefined customer clusters
• Membership to ad-hoc customer clusters
• Age
• Profession /schooling
• Household structure, wealth, revenue
• ZIP code
• Lifestyle
• ….
Customer choice

• Firms must understand

• Purchasing processes enacted by customers
• The influence cast by intermediaries (resellers,
installers, etc.)
• Consumers’ purchasing process is based on
• choices made individually / by the household
• influences external to the market (marketing action)
• influences internal to the market (imitative processes)
 Customer choice
• Consumers' purchasing process (and future experience of the product) is based on the interplay
between their own "needs" and "needs satisfaction" offered by the product
• There are multiple ways to frame and analyze this phenomenon
• Needs do not all have the same importance (Maslow's hierarchy… more a commonplace topic than
a scientifically grounded theory)
 Customer choice
• Consumers' purchasing process (and future experience of the product) is based on the interplay
between their own "needs" and "needs satisfaction" offered by the product
• There are multiple ways to frame and analyze this phenomenon
• Kano's model suggests needs do not all have the same importance, and evolve in time
("delighters" eventually become "order winners" and then "qualifiers")

(delighters)

How would you feel if the product had less?

(order winners)
Like it Expect Don't Can live Dislike
it care with
How Like it Attractive Performance
would
Expect it
you feel if (qualifiers)
the Don't care Indifferent Must-be
product Can live with
had
more? Dislike
 Customer choice
• Consumers' purchasing process (and future experience of the product) is based on the interplay
between their own "needs" and "needs satisfaction" offered by the product
• There are multiple ways to frame and analyze this phenomenon
• Needs can be studied at different depths, from a marketing or a design perspective
Design perspective

Marketing perspective

Technical solutions Generic needs

• Delighters Fundamental needs
Technical requirements
•Order winners • outcome = “job to be done”
Technical specifications •Qualifiers

“I want my razor to give me a deep shave”

"I want my razor to glide smoothly on my skin"
“Thre razor should have three blades” “I don’t want to see any hair growing in the evening”
"The razor should have a lubricated strip" “My current razor irritates my skin. I don’t like it”
 Customer choice
• Consumers' purchasing process (and future experience of the product) is based on the interplay
between their own "needs" and "needs satisfaction" offered by the product
• There are multiple ways to frame and analyze this phenomenon
• Needs can be expressed with Empathy Maps, trying to understand customers' lives and the
ensuing (expected) "gains" and (unwanted) "pains"…
 Customer choice
• Consumers' purchasing process (and future experience of the product) is based on the interplay
between their own "needs" and "needs satisfaction" offered by the product
• There are multiple ways to frame and analyze this phenomenon
• Needs can be expressed with Empathy Maps, understanding customers' lives, and trying to
understand the ensuing (expected) "gains" and (unwanted) "pains"… and then using Value
Proposition Maps to express how the product can cope with them

The customer
• Perceives the value proposition in
relation to the needs he feels
• Chooses and buys the product
• Uses the product and reassesses his
perception of product and needs

The firm
• Studies customer needs
• Designs the product to address them
 Customer choice
• Consumers' purchasing process (and future experience of the product) is based on the interplay
between their own "needs" and "needs satisfaction" offered by the product
• There are multiple ways to frame and analyze this phenomenon
• Customer choice is not directly based on the objective features of the product, but on a
perceptual process of product quality (intrinsically and relative to competition)→ lens model

Product
This is features Perception Preferences
objective and
has to do with Initial
“product need Choice
superiority”
Awareness
Advertising This is
Influence of subjective and Availability Actual
distribution has to do with market
Word-of-mouth “product
meaningfulness” share
Customer choice

• How do customers perceive products?

Perceptual space

Primary needs
(perceptual dimensions)

Secondary needs

Tertiary needs
Customer choice
• Tertiary needs
• A very long list of “elementary needs” (≠ technical features or parameters)
• Can be compiled by using research methods based on interactions between analysts
and customer samples (user-centered research) → attributed to personas
• Are nowadays being obtained also from semantic analysis of
• Internal IT systems (e.g. claims databases)
• Web sources (blogs, discussion lists, Virtual Customer Environments, etc.)
• Can be extended to complex influence networks (especially in the case of B2B)
 Customer choice
Methods for qualitative customer research
• Focus groups
• Groups with 8-12 people and a facilitator, discussing in a neutral environment
• Sample is not representative (wide spectrum)
• “Difficult to attract” customers can be under-represented
• Lead users play a key role
• The facilitator is a critical resource (must widen the discussion but not steer it)
• The analyst must be able to codify tertiary needs by taking part to the focus group and
by analyzing A/V recording
• High cost → groups are organized until not enough new needs arise
• Direct observation
• Similar to focus groups, but based on people using the product in an operational setting
• Scenario of use
• Similar to ”direct observation”, but participants (5-20) must act according to a script (e.g.
“Sunday morning at Mr. Black’s home”)
• Participants can work in groups, with the audience making comments on the scenes
being represented
 Customer choice
• User trials
• “Customers” (8-25) must use products or prototypes by performing a sequence
of planned activities
• Participants are interviewed after the session (no direct observation)
• Product-in-use
• Direct observation of the product being used in real settings by unsuspecting
people (data logging, A/V recording,…)
• Useful for getting hold of original or wrong “use modes”
• Customer diaries
• Participants use products at home and record their comments (paper forms,
online websites, A/V recording, etc.)
• Low cost, but information gathered is “shallow”
 Customer choice
• Consumer idealized design
• Customers are split in subgroups, design their “ideal product” and present it to
the others
• Design is usually supported by paper, lego bricks, clay, etc.
• Product personality profiling
• Within a focus group, customers must suppose the product is a person, and
describe personality, lifestyle, etc.
• Web of associations (proprietary method of Brandgenetics inc.)
• Participants are shown concepts, brands, products, and must associate them to
keywords (usually drawn from a list)
• Results are analyzed statistically
• Powerful method for associating brands to “values”
• Mood boards
• Participants must associate immages (e.g. making a collage from magazine
clippings) to products or brands
• The analysis is quite “deep”, but is difficult to interpret
Customer choice
• Online communities
• Branded sections of established online communities (e.g. Facebook)
• Branded but open communities (Nike+)
Four modes of active customer involvement in NPD
• Private online communities (Fuchs and Schreier 2011) These modes can be applied
• Private online panels differently to
• The entire user / installed base - The product
- Modules of the product (core vs. non-core)
• Actual usage and associated - Ancillary aspects of the product (e.g. advertising)
reports
• Analysing data generated by
users / products
• Listening to opinions expressed
online
• Observing general trends
occuring online (netnography)
• Communicating and looking at
reactions (from teasers to full-
blown ads)
• Explicitly involving customers
within the product development
process
 Customer choice
• Secondary needs (about 10)
• Tertiary needs are aggregated
• Aggregation can be done by experts (e.g. the product
development team)
• Similarity matrices
• A sample of individuals group tertiary needs (usually with cards)
• The number of times in which need i is grouped with need i’ defines a
x
similarity indicator ii’
• hierarchical clustering allows to define groupings among secondary needs
Customer choice

• Primary needs (principal perceptual dimensions)

• Primary needs are a “tacit” or “latent” aggregation of
correlated secondary needs
• Primary needs should ideally be 2 or 3
• Primary needs cannot be found directly, but must be
elicited through
• Quantitative market research (surveys)
• Multivariate data analysis (factor analysis or
multidimensional scaling) on the data set coming from the
survey
 Product positioning

• Customers can be represented on perceptual maps

14

Can you see evident

clusters?
12

Is cluster memberhsip associated

10
to descriptive variables to a
statistically significant degree?

8
DOM_6A

1,00
Is the position on the
perceptual map associated to
AAA

6 ,00
4 6 8 10 12 14 16 descriptive variables to a
statistically significant degree?
BBB
 Product positioning

• Products too can be represented on perceptual maps

14

12

10

8
DOM_6A

1,00
AAA

6 ,00
4 6 8 10 12 14 16

BBB
Product positioning
• Products can also be represented on “value maps”
14

12

10

8
DOM_6A
AAA/$

1,00
AAA

6 ,00
4 6 8 10 12 14 16

BBB
BBB/$
 Market research
• "Analytics" on existing "big data" are complementing and
progressively substituting customer surveys to gain
quantitative information on customer preferences
• Survey design and administration – the key problem is
cost… which depends on
• the breadth of the questionnaire
• the sample size
• the administration medium
• Personal interviews
• Telephone interviews
• Mail, fax
• E-mail or web-based
• the size of the pilot sample
Market research

• Survey design and aministration – representativeness of

the sample … depends on
• Depends on administration strategy
• Depends on biased response rates and missing data (function
of the classes of respondents, on the questionnaire, on
incentives)
• Must be verified ex-post at both levels (questionnaire and
individual variable)
Market research

• Survey design and administration – the need to

be focused
• For the sake of methodological rigour (set hypothesis,
define the approach to be used for analysis → then
use the approach to verify the hypothesis), opposed
to data fishing
• In order to design questionnaires that are
• Parsimonious
• Where questions have the correct meaning
• Where variables are appropriate to the desired statistical
analysis method
Market research

• Survey design and administration – questionnaire

size
• Questionnaires must cover all relevant aspects
• Must lead to sufficiently large and non-biased
samples (as a rule of thumb, multivariate statistics
require approx. 10 respondents per variable to be
analyzed simultanesouly)
• The larger the questionnaire, the larger the minimum
sample size … but the lower the response rate → cost
is more than proportional to the size of the
questionnaire
Market research

• Survey design and administration –

Questionnaire structure
• Clear, divided in sections, must be easy and “fun” to
reply to
• Easy questions at first, for kick-starting the response
• Avoid repetitions
• Must follow the respondent’s logic, and not the
analyst’s
• Avoid jumps
Market research

• Survey design and administration - Questions

• Must be formulated in a clear and non-ambiguous way
• Find the right tradeoff between using technical terms,
explanatory notes, or accepting some degree of ambiguity
• Questions must be “neutral”
• Survey design and administration – Responses
• Must be clearly appropriate to the statistical methods to be
used
• Must be easy to fill in
• Questions can be “open” or “closed”
Market research

• “Membership” or “choice”
• “Open” replies (and possibly ex-post coding)
• Selection on “closed” responses (always include an open
“other ____” category)
• Choose k out of m items
 Market research
• When expressing “agreement” or “importance”
• Likert / verbal scales (odd or even levels?)
• Numerical scales (semantic differential)
• Item ranking
• Scoring (1-10 or 1-100)
• Spreading a score of 100 over many items
• Selecting k of m items
• Take care
• Scales are ordinal in nature
• Different conventional meanings can be attached to values (e.g.
5/10 vs. 6/10)
• Should you normalize w.r.t. the respondent’s average scores?
Market research

• Questionnaires allow you to determine

• The importance given by each customer to secondary needs
• The perceived performance of each product w.r.t. each
secondary need
• Multivariate analysis leads from the raw data collected
through the questionnaire to the (latent) perceptual
dimensions
• Perceptual maps allow you to locate
• Customers → define segments and estimate their size
• Products (currrent & future) of the firm and its competitors →
strategic positioning of the product
 Market research
• Survey-based methods require a representative sample of a
finite population
• Respondents are chosen at random (probability sampling) → unbiased
• Sample size is sufficiently large → confidence intervals are not too wide
• Probability sampling is the golden standard
• Random sampling (e.g., random phone numbers)
• Systematic sampling (e.g., order the population and pick one every N)
• Stratified sampling. Define a number of relevant strata and perform random
sampling on each stratum in order to obtain a sufficient sample size per stratum
• Nonprobability sampling
• Convenience sampling (select an easy-to-reach sample)
• Judgement sampling (e.g., assume city X is representative of the nation)
• Quota sampling. Define strata and use convenience or judgement sampling in
each stratum
• Snowball sampling (ask respondents to involve other respondents)
 Market research
• Stratified sampling
• The population is divided in subgroups (strata) so that
each member belongs to one and one only stratum
• Sampling is carried out randomly in each stratum.
• Proportional sampling - Sample size in each stratum is
proportional to stratum size → unweighted results
• Nonproportional sampling – Sample size is not proportional,
either because of difficulties in reaching the target or because
oversampling is performed in strata with high variance →
weighted results
• Optimal sampling is a nonproportional sampling where sample
size in each stratum is proportional to stratum variance and
inversely proportional to the cost of reaching each element in
the stratum
 Market research
• Sample size – case of proportions (i.e. percentage of
the population that exhibits a feature)
• The estimator is 𝑝Ƹ = 𝑋Τ𝑛, where X is the observed
number of positive observations and n is sample size
• The estimator has a Bernoulli distribution whose
variance is maximum at p=0.5 and equal to 0.25/n
• If n is large, it can be approximated to a normal. The
confidence interval at  significance level is 𝑝Ƹ ±
𝑍𝛼 0.25Τ𝑛
• If  = 95%, Z≈ 2 and the error will be ±2 0.25Τ𝑛 =
± 1Τ𝑛 so that n = 1/E2 with E the percentage error
 Market research
• Sample size – case of means
• The estimator is the arithmetic mean 𝑥ҧ
• The confidence interval of the mean estimate at 
significance level is 𝑥ҧ ± 𝑍𝛼 𝜎Τ 𝑛
• If  = 95%, Z≈ 2 and the error will be ±2 𝜎Τ 𝑛
• Given the desired error  E, it is straightforward to derive
n, given an estimate of 
Elements of multivariate statistics for
market research
Outline of the presentation

• Factor Analysis
• Hierarchical clustering

The approach is very practical ("how to") and

definitely not theoretical
Factor Analysis – Possible objectives

• Objective 1 – Finding “structure” in the dataset

• Group similar (= correlated) variables → identify latent
variables
• Group similar individuals → identify clusters (but cluster
analysis techniques are more commonly used to this purpose)
• Objective 2 – Data reduction
• Use latent variables as syntethic variables
Factor Analysis - Planning
• Identify m random variables
Observations
 x1   x11 ... x1n 
 
Variables (e.g., "how
x =   the dataset is x1 ,..., x n =  ... ...  important is secondary
need xyz to you?"
 x m   x m1 ... x mn 

• You need a sample of size

• n > 100 (at the limit 50)
• n > 10 m (at the limit 5 m)
 These variables are
Factor Analysis - Planning correlated to one
another… do they
exhibit a same and
• What are we looking for? higher-level meaning?
Observations

1 4 2 4 2 5 3 5 1 2 3 4 1 3 4
2 5 1 3 1 4 2 4 2 2 4 4 2 2 5
 Variables (e.g., "how
1 5 2 4 1 4 3 5 1 1 3 5 1 3 4
x=  important is secondary
4 1 3 5 5 5 4 4 2 1 4 3 4 5 3 need xyz to you?"
5 2 3 5 4 5 4 5 1 2 5 3 5 5 3
 
5 1 2 4 5 4 5 5 1 1 5 3 4 4 2
Same for these variables

And, by the way…. the two

groups of variable do not appear
to be correlated with one another
Factor Analysis – Initial hypothesis
• You must check that the   11 ...  1m 
covariance (or correlation)  =  ... ... 
matrices exhibit “structure”  m1 ...  mm 
(some correlations are high can be estimated by
and other ones low)
S=
1
n −1

n
(
j =1
)(
xj − x xj − x)T

where
1 n
x =  j =1 x j
n
 ij Covariance suffers from
ij = scale effects.
 ii jj Correlation matrices
work better
 Factor Analysis - Planning
Observations
1 4 2 4 2 5 3 5 1 2 3 4 1 3 4
• What are we 2
 5 1 3 1 4 2 4 2 2 4 4 2 2 5
looking for? 1
x=
5 2 4 1 4 3 5 1 1 3 5 1 3 4
 Variables
4 1 3 5 5 5 4 4 2 1 4 3 4 5 3
5 2 3 5 4 5 4 5 1 2 5 3 5 5 3
 
5 1 2 4 5 4 5 5 1 1 5 3 4 4 2
 Factor Analysis – Initial hypothesis
• In order to verify the existence
of “structure”
• Check whether any correlation is
> 0.3
• Bartlett’s sphericity test (H0 = no
correlation is significantly > 0).
The test is sensitive to n
• MSA – Measure of Sampling Adequacy  [0,1], should be > 0.7
• Compute global MSA
• Compute MSA for each variable, and discard the ones with lower MSA until the
overall value of MSA is all right
Factor Analysis - Execution
• Choose a method for FA. The differences lie in the algorithms and in
the variance being considered
Between variables → the higher, the
Common greater the correlation is
To each variable → the higher, the
Specific
lower the correlation is
Error Residual

Approach Variance being Use

considered
Principal Simple All variance When you know that common
Components algorithm variance is large
Analysis
Common Statistically Common variance When you don’t know the
Factor rigorous only proportion between variances
Analysis approach When you do exploratory work
 Factor Analysis - PCA
• You have a vector x of stochastic variables, and matrix S represents covariances
• Calculate a first linear combination of this vector x, which you will call y1
(which is a scalar stochastic variable)
• Determine coefficients of this linear combination so that y1 exhibits maximum
variance
• Remember: when you sum correlated

( )
variables, you get a larger variance than the sum
max Var ( y1 ) = Var a1 x = a1 S a1
T T
of variances!

s.t. •In other words, by trying to maximize variance

of y1, you will allocate greater weightings ai to
a1 a1 = 1
T
the variables that vary together (which are
highly correlated) and lower weightings to the
other variables (which are less correlated)

You want to avoid • If variance of the resulting y1 is k% of total

trivial responses with variance, you can say that the new variable y1
a1= m explains k% of total variance
 Factor Analysis - PCA
• Now, calculate a second linear combination of x, which you will call y2,
so that it is orthogonal (i.e. uncorrelated) to y1, so that you maximize
its variance … and so on The variance of each resulting variable y will be a
i

( )
% of total variance.
max Var ( y2 ) = Var a x = a S a 2
T T
2 2
Variable yi explains that same % of total variance
s.t.
Variables y1, y2,…yk collectively explain a % of
a a2 = 1
T
2 total variance

a 2 a1 = 0
T

( )
max Var ( yi ) = Var a i x = a i S a i
T T

s.t.
ai ai = 1
T

a i a i ' = 0 i '  i
T
 1) If A is square, x and  are an eigenvecto r and eigenvalue if A x =  x

Factor Analysis (A ) =   ,
- PCA
2) trace i i A = i i
3) x i x i ' = 0, i  i '
T

• What is
4) if x i x i = 1 (i.e., we normalize eigenvecto rs), x i A x i = i
T T
happening in
matricial terms? •a1 is the eigenvector of S ( )
max Var ( yi ) = Var a i x = a i S a i
T T

• The optimization that corresponds to its

s.t.
problems are greatest eigenvalue
ai ai = 1
T
same as working •ai is the eigenvector of S
with eigenvalues a i a i ' = 0 i '  i
T
that corresponds to its i-th
and eigenvectors eigenvalue (in decreasing
order)

i=1 i = trace(S ) S a1 = 1 a1
m

• The total variance of the m and

Var ( y1 ) = a1 S a1 = 1
principal components is given by T

 i '
i'
• … and the first i’ principal components i =1
explain a fraction of the total variance trace(S )
 Factor Analysis - PCA
• Now, let's make a step forward
1 0 0 
Let us have  =  0 ... 0  and A = a1...a m 
 0 0 m 
you can show that S = A A
T

If we normalize eigenvectors according to the eigenvalue, a~ i a~ i = i (not to 1)

T

~ ~T
you have a~ i = i a i and S = A A
~
 
where the matrix of re - normalized eigenvectors A = a~1 ...a~ m is called the matrix of factor loadings
(tells you the incidence of the row variable on the column factor)

• The matrix of factor loadings allows you to work on the relationship

between variables and underlying factors. For instance…
We can determine the value of individual j on factor i
y = a~ x
T
ij i j
Factor Analysis - PCA
 Factor Analysis - CFA
• PCA is only one way to determine factors underlying a set of variables
• CFA follows another approach. We hypothesize an ex ante model based
on k<m latent variables
• The model says that the m original and X =  f +u Residual
observed variables X can be estimated
as a linear combination of the k Factors
unobserved factors f, with an error u (latent variables)
• You can use a variety of estimation
methods (usual objective is to make Loadings
the correlation matrix of the estimated
X as close as possible to the original  11 1k 
= 
one… in other words you try to
minimize the correlation matrix of u)  
• You run the algorithm multiple times, m1 mk 
with k = 1, 2,…m
Factor Analysis – How many factors?
• % of total variance explained (60%?)
• Scree plot (look for the “knee”)
• “Latent root” criterion (keep factors with i>1… a factor with eigenvalue
< 1 explains less variance than a single variable)
• A number defined a priori (e.g. 2 or 3)
 Factor Analysis – Factor rotation
• Warning! Factor loadings do not help understand structure in the dataset,
until they are rotated in such a way to facilitate interpretation
• Factor loadings allow you plot the variables on the factor space
• A clearer solution can be
found, if you are able to
rotate the axis so that
• → factor loadings become
closer to 0 or to 1
• → the way each factor is
assigned to a group of
variables (or each variable
to a factor) becomes more
evident
• What if you can rotate and
get rid of orthogonality?
 Factor Analysis – Factor Rotation
• Orthogonal rotation
• VARIMAX (simplifies columns →
which variables are tied to a factor)
• QUARTIMAX (simplifies rows → to
which factor should you tie the
variable)
• EQUIMAX (hybrid)
 Factor Analysis – Factor Rotation
• Oblique rotation (OBLIMIN),
creates a “neat” result, but you
lose orthogonality (i.e., factors
become correlated… but is that a
problem?)
• You usually consider loadings >
0.5 (i.e., variance of the variable
is explained at least at 25% level
by the latent variable)
• You can sometimes compute the
statistical significance of the
loading
 Factor Analysis – Creation of synthetic
variables
• Surrogate variables (the latent variable is represented by the single original
variable that loads most heavily on it)
• Summated scales on the variables that “belong” to the factor (without any
weights). The scale must be validated with reliability measures (Cronbach
 > 0.6-0.7)
• Factor scores (original variables are weighted according to the factor
loading)
Factor Analysis and perceptual mapping
• Questionnaires provide the
importance of each secondary need
i to subject j, xij
• When factor loadings have been z jk = iik xij
computed, ik we derive the
importance of each primary need k
to subject j, zjk

• If we have asked each subject j to

evaluate product l on secondary 1 
need i, qijl we can derive the wkl = iik   j qijl 
positioning of product l in the n 
perceptual space, wkl
• More complex models allow to
obtain everything by simply starting
from qijl
Hierarchical clustering – Distances
• We know xij = value scored by
 (x − x ) 
1
individual j on variable i euclidean d jj ' = ij ij '
2 2
i
• We compute a distance
city block d jj ' = i xij − xij '
(dissimilarity) matrix between
individuals
 x − x 
1
r r
Minkowsky ( r ) d jj ' =
• Variables are generally i ij ij '

normalized on a (0,1) interval or Norm max d jj ' = max xij − xij '
i
according to average and std.
deviation, so to avoid scale
effects
• The number of variables should
also be balanced (avoid
unwanted weightings to arise
because of multicollinearity)
 Hierarchical clustering - Aggregation
• Once an aggregation criterion has been chosen, at each step
you find the “closest” clusters, you merge them, and you
recompute distances, etc.
 Hierarchical clustering - Aggregation
• Once an aggregation criterion has been chosen, at each step
you find the “closest” clusters, you merge them, and you
recompute distances, etc.
 Hierarchical clustering – Aggregation
criteria
• Single linkage (join the two clusters whose
closest members are closest → “chain-like”
clusters)
• Complete linkage (join the two clusters
whose furthest members are closest →
“spherical” clusters
• Average linkage (join the two clusters that
exhibit the smallest average distance
between all members → clusters of similar
size)
• Ward (join the two clusters that exhibit the
smallest quadratical distance between each
member → clusters with smilar number of
members)
• Centroid (join the two clusters whose center
of gravity is closest → tricky and difficult to
interpret, because centers of gravity change
at each step)
Pricing, positioning, forecasting
Product positioning

• Balance between performance and price

• Pricing is a difficult problem (does price still exist?)

Reservation price

Price

Profit
Contribution margin
Fixed costs (how are they computed?)

Variable cost
 Price = VC * (1+ fixed margin)

Product positioning
Price = VC * (1+ product-dependent margin)

Price set so that Marginal Revenue =

Marginal Cost (or similar methods)

Price defined by using competitors' prices as

benchmarks (considering product quality)
Price defined by estimating the value
generated, from the customers' perspective
Price defined with reference to customers'
(or producers') breakeven threshold

Price defined with reference to

customers' (or producers') IRR

Price defined by using competitors' prices as

benchmarks (indifferentiated product)

Internal pricing between BUs of a same firm, balancing compliance to norms,

market prices, internal incentives and opportunistic usages of the mechanism
 Pricing
• Demand-based pricing methods

Price Sensitivity Meter (van Westendorp)

100
Monadic Price Concept Test 90
Indifference
price point

Cumulative % of respondents
Sales 80 (normal
price)
in test 70

event 60
50
Marginal point of
40 cheapness (lower bound Marginal point of
of acceptable price) expensiveness
price 30 (upper bound of
acceptable price)
20
10 Optimal price point
0
0
00
00
00
00
00
00
00
00

10 0
11 0
12 0
13 0
14 0
15 0
16 0
17 0
18 0
19 0
20 0
0
0
00
00
00
00
00
00
00
00
00
00
00
10
20
30
40
50
60
70
80
90 Inelastic range Price

Too cheap Bargain Expensive Too expensive

 Product positioning
Segmentation criteria for B2B
• General criteria for Industry
segmentation
Firm size
• Product features
• Quantity purchased Impact on cost structure
• Previous purchasing Quantity purchased
behaviour Purchasing channel / procedures
• Purchasing channel
• Geography Segmentation criteria for B2C
• Price sensitivity Demographics
Life-style
Brand preferences and brand switching
behaviour
Perception of brand
Relationship with advertising channels
 Competitive landscape theory
• Levinthal (1997) and other authors depict customer preferences as Willingness-To-
Pay (or utility?) as a function of product attributes (or primary needs satisfaction?)
• Each customer (or each customer segment exhibiting similar utility function?) is
characterized by his own landscape
• Firms can be represented as actors who adapt their offering, moving towards the
maxima on the landscape either by local search, or by trying to model the landscape
→ this does not disrupt the landscape to competitors

B Firm alpha moves from

A to B in order to reach
a point where
A utility/WTP is higher
 Competitive landscape theory
• However, firms, can also act on the landscape, altering customer preferences (Vinokurova,
2019), which also potentially disrupts competitors. Which of the two (i.e., product or
landscape) is more "malleable"?
• This can be done by
• Transforming dimensions (i.e., given dimension X, you can act on the way with which customers
react to change on X, or on the weight they attribute to X…. at the limit you can even remove X)
→ "drive safely" flattens customers' reaction to performance in cars
• Adding a previously non-existent dimension, or enhancing a poorly appreciated dimension one
→ introduction of "infotainment contents" in cars
Firm Beta alters
customers' utility
function. A is
already close to
the maximum
B
A
A
 Product positioning
• From an economics point of view, positioning implies
differentiation
• Horizontal: by varying product features, utility increases for some
customers and decreases for others (e.g. taste of food products)
• Vertical: by varying product features, utility increases or decreases
in the same sense for all customers, albeit with varying degrees
(e.g. fuel consumption of a car)
 Product positioning
• Horizontal differentiation
• Allows you to get close to the maximum utility of a given segment → you can ask
for a higher price
• Redistribution of competitors may reduce competitive pressure (Makadok and
Ross, 2013)
• If horizontal differentiation allows to increase sales you also can have
• Economies of scale, scope and learning
• Overcome minimum efficient size thresholds
• If volume does not change, the opposite holds, and the organization becomes
more complex
Reservation Segment 1 Without
price differentiation

Segment 2

Parameter x
With differentiation
 Product positioning
• Vertical differentiation
• Price discrimination allows higher profits
• Redistribution of competitors may reduce competitive pressure
(Makadok and Ross, 2013)
• If vertical differentiation allows growth in volume you also can
have
• Economies of scale, scope and learning
• Overcome minimum efficient size thresholds
• If volume does not change, the opposite holds, and the
organization becomes more complex
• Risk of “cannibalization”
 Product positioning

• If utility curves for two segments • If utility curves do not cross, you
cross, you don’t have cannibilization have cannibalization
(sensitivity to quality is different for • A is better off by purchasing the “low
the two segments quality” product
• A does not buy the “low quality”
product
• B does not buy the “high quality”
product
Reservation price Reservation price
segment A Price for segment A
Price for
segment A
segment A
Price for segment B
segment B
segment B
Price for
segment B
low high quality low high quality
 Product positioning
• Avoiding cannibalization
• Reduce price of high quality product →A gets the same surplus as with B (profits suffer)
• Make the sensitivity curve for A steeper (marketing)
• Degrade the low quality product (locate it to the left of the intersection between lines
• Differentiate A and B horizontally → give B features that B likes and A dislikes (e.g. a
consumer product with predefined functions and no flexibility for professional use)
• Sell either A or B (niche strategy)
• Launch A before B
• The optimal choice is fairly complex (Moorthy e Png 2002, Krishnan e
Zhu 2006) and depends on
• The steepness of the utility curves
• The relative size of segments A and B
• Development and variable costs for the two products
• The interest rate (degree of impatience) of the firm and of the market
Demand forecasting  exogeneous?

• The most commonly used model Depends on product features

Demand for product =
Depends on advertising
market demand x
x potential market share x Depends on distribution
x awareness x
x availability
• How can you determine the four
parameters?

TAM = market demand

SAM = TAM x awareness x availability
SOM = SAM x potential market share
Demand forecasting

• Market demand
• In stationary markets you use time series analysis or
regressions (possibly correlating demand with exogeneous
variables describing the economy, demographics, etc.)
• When diffusion dynamics are important diffusion models must
be used
 Demand forecasting
• The Bass model (1969) is the basic diffusion
model
• Underlying hypothesis: Cumulated sales
• Monopoly or market-level demand 1800000
Stock of • Durable and innovative good (no substitutes or 1600000
1400000

Sales [units]
1200000
adopters at complements)) 1000000
800000
time t • Constant marketing actions 600000
400000
Innovative adoption 200000

• Binary adoption process coefficient →

0

0
2
4
6
8
10
12
14
16
18
20
22
24
26
28
Time bucket
advertising
Innovative scenario Imitative scenario

. Sales

= pM − N (t ) + q M − N (t )
dN (t ) N (t )
n(t ) = 250000

200000

Sales [units]
dt M 150000

100000

Sales at 50000

Final number of Imitative adoption 0

time t
adopters coefficient → WOM and

0
2
4
6
8
10
12
14
16
18
20
22
24
26
28
Time bucket
network externalities Innovative scenario Imitative scenario
Demand forecasting
The Bass model can be interpreted as a survival model (i.e.
survival in the state of non-adopter)
The model is defined by a hazard function (t)
(t) is the probability of adopting the product in (t, t+dt), under the
condition of not having adopted up to t
If f(t) is the probability density function of the random variable “time to
adoption” and F(t) is its distribution function (= probability of having
adopted before t),
f (t )
 (t ) =
And therefore
1 − F (t )

 N (t ) 
M − N (t ) =  (t )M − N (t ),  (t ) =  p + qF (t )
.
n(t ) =  p + q 
 M 
Demand forecasting

• The Bass differential equation allows a closed-form solution for n(t)

e N(t)
( p + q)2 e − ( p + q ) t 1 − e −( p + q )t
n (t ) = M 2
N (t ) = M
  q −( p + q ) t
q
p 1 + e −( p + q ) t  1+ e
 p  p
If q=0, diffusion is innovative If p=0, diffusion is imitative
(Fourt e Woodlock, 1960) (Mansfield, 1961)

n(t ) = Me − pt M
N (t ) =
1 + e −qt
(
N (t ) = M 1 − e− pt )
Demand forecasting
Sales peak is characterized as follows
n(t ) 1 q
t* | = 0, t* = ln  Position in
t p+q  p time

M ( p + q)
2
n (t *) =
Sales peak
4q

1 p  M
N (t *) = M  −   Penetration
 2 2q  2 level
Demand forecasting

• Parameters can be estimated by using autoregressive forms

nt = N t +1 − N t = p(M − N t ) + (M − Nt )Nt
q
M

N t +1 = pM + N t (1 − p + q ) −
q 2 Multicollinearity
Nt
M !!!
If q=0 If p=0
N t +1 = N t (1 + q ) −
q 2
Nt +1 = pM − Nt (1 − p ) Nt
M
Demand forecasting

• A number of extensions to the Bass model are available

• Horsky (1990), where the market potential varies in time as a
function of
• The maximum potential M
• The price of the good pr(t)
• The utility of the good, K
• Available income within the population (average w(t) and variance
var[w(t)])  
 
− N (t ) p + qN (t )
M
n(t ) =  K + w(t ) − pr (t )
−
 
1+ e var w(t )  
Demand forecasting

• Generalized Bass model (1994), where instantaneous sales are

moderated by relative variation of
• Price pr(t)
• Marketing effort A(t)

Becomes identical to

n(t ) = M − N (t ) p + qN (t )x(t )

the normal Bass model
when pr e A exhibit
constant percent

dpr(t ) dA(t )
variations (→ negative
x(t ) = 1+  p + A exponential), which is
pr(t ) A(t ) often the case for hi-
tech goods
 Demand forecasting
• Diffusion models with substitution and additional sales (parameter
estimation requires to disaggregate the two!)
Stock of
l
st = nt + rt +  vit Addition sales of the i-th customers that
Total sales own i goods in t
i =1
good in t
in t
vit = hi (t ,Vit )(iVi −1,t − Vi ,t )
Adoption sales Substitution
in t sales in t
Hazard function Fraction of
t for the purchase customers
rt =  (Rt −t ' − Rt +1−t ' )st ' of the i-th good willing to buy
t '=1 more than i
goods
Probability that the good is Probability of a good purchased in t’ to
still usable beyond t-t’ break down between t and t+1
 Demand forecasting
• A rough approximation of repeat
sales is
rs(t ) = N (t )
 
 −( p + q ) t 
• Total sales therefore amount to + 2 −( p + q ) t
−
ts(t ) = n(t ) + rs(t ) = M  
( p q ) e 1 e
+ 
  q −( p + q ) t 
2
q −( p + q ) t 
 p 1 + e  1 + e 
• It can be shown that cumulated p
  p  
repeat sales are given by

1 − e−( p+q )t M
RS (t ) =  rs(t )dt =  M
q
dt =
q

ln ( pe( p + q ) t + q ) − pt − ln( p + q ) 
1 + e −( p + q ) t
p

TS (t ) = N (t ) + RS (t ) =
• Therefore, total cumulative sales
amount to  
 1 − e −( p + q ) t  
= M
q
 
+ ln ( pe( p + q ) t + q ) − pt − ln( p + q ) 
1 + e − ( p + q ) t q 
 p 
 Demand forecasting
• In services, recurring revenues (e.g., monthly subscription fees) coming
from a market that is undergoing diffusion are similar to repeat sales

• With recurring revenue, instant dollar sales are defined by Average

Revenue Per User (ARPU) per time unit
S (t ) = ARPU N (t )

• Cumulative sales up to t therefore amount to

1 − e−( p+q )t
RS (t ) =  S (t )dt =  ARPU M
q
dt =
ARPU M
q
 
ln ( pe( p + q ) t + q ) − pt − ln( p + q )
1 + e −( p + q ) t
p

• Churn (% of customers who abandon the service) is not included!

 Demand forecasting

• Extension of the Bass model to complementary and substitute goods

n1 (t ) = p1 M 1 − N1 (t ) + q1 M 1 − N1 (t ) + r21 M 1 − N1 (t )
N1 (t ) N 2 (t )
M1 M2

n2 (t ) = p2 M 2 − N 2 (t ) + q2 M 2 − N2 (t )+ r12 M 2 − N 2 (t )
N2 (t ) N1 (t )
M2 M1

Indipendent Cross-good diffusion

diffusion effect
Demand forecasting

• A simplification of the extension to two-sided

markets (Chun and Hahn 2008)
• If there is a reciprocal complementarity with another
product

 N (t ) NM(t ) 
n(t ) = M − N (t )  p + q e 
 M 
 Demand forecasting
• Recent approaches work better on initial data by relating adoption
decisions to consumer utility (Decker and Yukawa 2010)
WOM + network
Nt-1 Expected Adoption occurs in t if
externalities utility of E(Ut) > E(Ut+1)
purchasing
Price changes Quality-adjusted in t, E(Ut)
price in t, Pt Purchasing
probability
Expected in t
Forward- Expected quality- utility of Sales in t, nt
looking adjusted price in purchasing in
behavior t+1, E(Pt+1) t+1, E(Ut+1) Remaining
market potential
M-Nt-1
If ut=0+1ln(Nt-1)+t   is the discount factor
and 2 captures the forward looking behavior
M − N t −1
nt =
1 + e −( 0 + 1 ln ( Nt−1 )+  2 ( Pt −Pt+1 ))
 Demand forecasting
• For B2B markets, demand is often driven by direct sales effort
• The sales process needs to be modeled, estimating probabilities and effort and
computing the cost of one sale
• Advertising has the effect of improving the performance of the process
• To date, there are no models bringing together direct sales, advertising and word-of-
mouth

1° call p1 2° call p2 3° call p3 4° call p4

(general (in-depth (negotiation) (closing)
presentation) business case)
e1 e2 e3 e4
• Diffusion of digital services is strikingly similar to direct sales, being driven by marketing
effort (→ metrics such as cost per impression, customer acquisition cost, etc.)
Demand forecasting

• Market share models are usually based on

“Kotler’s theorem” (market share si depends on
the “attractiveness” Ai of all competing goods)
• Axiom 1: Ai  0, i Ai > 0
• Axiom 2: Ai = 0 → si = 0
• Axiom 3: Ai = Aj → si = sj
• Axiom 4: if Ai := Ai + , sj  is a function of , not of i

Ai
si =
 j
Ai
Demand forecasting

• Market share models

• Technical features of the product must be associated to its
“attractiveness”
• Historical sales can help…
e.g. MNL, Multinomial Logit
 
Ai = exp i +   k X ki +  i 
 k 
MCI, Multinomial Competitive Interaction

Ai = exp( i ) X ki k  i
k
 Demand forecasting
Gather data from Estimate
products si, Xki implicit
competing in the attractiveness
market
Ai
Estimate
parameters
for MCI or
MNL models
i, k.
Create future X*ki Forecast A*i Forecast s*i
scenarios for attractiveness market shares
products competing for new for new
in the market products products
 Demand forecasting
• Models differ with respect to elasticity of si with respect to
parameters Xki
 k X ki si si
• Linear esi = esi =
si X ki X ki
• Multiplicative esi = k • Elasticity should →0 when si →1
• Elasticity should → 0 when Xki → ∞
• Exponential esi =  k X ki

• MCI esi =  k (1 − si )
elasticity
MNL
• MNL esi =  k (1 − si )X ki MCI

Xki
Demand forecasting

• In market share models price can be considered

in many ways
• Neglect it (highly competitive industries →
commodity products)
• As one of the parameters Xki
• By operating at the level of a segment, where price is
fairly fixed (e.g. 100-120 € products)
• For highly innovative products (100% market share,
and difficulties in estimating diffusion) you can use
• Information coming from Monadic Price or Van Westendorp
methods
• Purchasing intent at a given price (e.g. 0.4 x “% definitively
would buy” + 0.2 x “% I would probably buy”)
Demand forecasting

• All these models suffer from high uncertainty in parameter

estimation
→ “the only certainty of a forecast is that it will be proved wrong”
• Demand forecasting in the long term is difficult and dangerous
• Models must be continuously calibrated
• It is advisable to express forecasts as a confidence interval
• Statistical elaboration
• Scenario analysis
• Monte-Carlo simulation