Risk Measures based on loss distribution

Roberta Pappadà

March 26, 2022

Contact: rpappada@[Link] Data Science for Insurance

 Table of Contents

1 Risk Measures
Introduction
Value-at-Risk
VaR: Examples
Expected Shortfall
ES: Examples
ES vs VaR

2 VaR and ES: Properties

Non-subadditivity
Coherent risk measures

3 Other Risk Measures Based on Loss Distributions

Risk Measures based on loss distribution March 26, 2022 2

 Table of Contents

1 Risk Measures
Introduction
Value-at-Risk
VaR: Examples
Expected Shortfall
ES: Examples
ES vs VaR

2 VaR and ES: Properties

Non-subadditivity
Coherent risk measures

3 Other Risk Measures Based on Loss Distributions

Risk Measures based on loss distribution March 26, 2022 3

 Table of Contents
Introduction

1 Risk Measures
Introduction
Value-at-Risk
VaR: Examples
Expected Shortfall
ES: Examples
ES vs VaR

2 VaR and ES: Properties

Non-subadditivity
Coherent risk measures

3 Other Risk Measures Based on Loss Distributions

Risk Measures based on loss distribution March 26, 2022 4

 Risk Measurement
Introduction

Risk measures associates a financial position with a real number, a

statistical quantity, describing the conditional or unconditional loss
distribution of the portfolio over some predetermined horizon
in the unconditional approach we denote the df of the loss L = Lt+1
simply by FL
Risk measures attempt to quantify the amount of assets that an
insurer needs to retain to meet obligations
it is natural to base a measure of risk on the right tail of the loss
distribution (e.g. VaR, ES)

Risk Measures based on loss distribution March 26, 2022 5

 Purposes
Introduction

Computation of risk measures is crucial for many insurance applications:

to determine a measure of riskiness of insured claims
to compute the amount of capital needed as a buffer against
(unexpected) future losses to satisfy a regulator (risk capital)
as a tool in financial risk management

Risk Measures based on loss distribution March 26, 2022 6

 Pros and cons
Introduction

(Pros) The quantification of risk associated with a given loss distribution

has some advantages:
the concept of a loss distribution allows for aggregation at different
levels
if estimated properly, the loss distribution may provide an accurate
picture of the risk in a portfolio
loss distributions can be compared across portfolios
(Cons) Two major issues arise when working with loss distributions
estimates of the loss distribution are based on past data
the assumption of normality is unrealistic in many situations, hence
alternative statistical models are often needed
risk measures based on the loss distribution should be complemented
by information from hypothetical scenarios
Risk Measures based on loss distribution March 26, 2022 7
 Table of Contents
Value-at-Risk

1 Risk Measures
Introduction
Value-at-Risk
VaR: Examples
Expected Shortfall
ES: Examples
ES vs VaR

2 VaR and ES: Properties

Non-subadditivity
Coherent risk measures

3 Other Risk Measures Based on Loss Distributions

Risk Measures based on loss distribution March 26, 2022 8

 Preliminaries
Value-at-Risk

We want to define a statistic based on FL which measures the severity of

the risk of holding our portfolio over a fixed time horizon ∆t

Risk Measures based on loss distribution March 26, 2022 9

 Preliminaries
Value-at-Risk

We want to define a statistic based on FL which measures the severity of

the risk of holding our portfolio over a fixed time horizon ∆t
A reasonable option is to consider the maximum possible loss, given by
inf{` ∈ R : FL (`) = 1}
::::
Value-at-Risk (VaR) can be viewed as an extension of maximum loss,
which takes into account the probability information in FL , by means of a
given confidence level [Jorion (2007)]
::::

Risk Measures based on loss distribution March 26, 2022 9

 Preliminaries
Value-at-Risk

We want to define a statistic based on FL which measures the severity of

the risk of holding our portfolio over a fixed time horizon ∆t
A reasonable option is to consider the maximum possible loss, given by
inf{` ∈ R : FL (`) = 1}
::::
Value-at-Risk (VaR) can be viewed as an extension of maximum loss,
which takes into account the probability information in FL , by means of a
given confidence level [Jorion (2007)]
::::
It was introduced by JPMorgan in the first version of its RiskMetrics
system

Risk Measures based on loss distribution March 26, 2022 9

 Preliminaries
Value-at-Risk

We want to define a statistic based on FL which measures the severity of

the risk of holding our portfolio over a fixed time horizon ∆t
A reasonable option is to consider the maximum possible loss, given by
inf{` ∈ R : FL (`) = 1}
::::
Value-at-Risk (VaR) can be viewed as an extension of maximum loss,
which takes into account the probability information in FL , by means of a
given confidence level [Jorion (2007)]
::::
It was introduced by JPMorgan in the first version of its RiskMetrics
system
::::
Computation of VaR involves quantiles of the loss distribution, hence we
recall definitions of the generalized inverse and quantile function

Risk Measures based on loss distribution March 26, 2022 9

 Generalized inverses and quantiles
Value-at-Risk

Let F be a df on R.
(i.) The generalized inverse of F

F ← (y ) := inf{x ∈ R : F (x) ≥ y }

is called the quantile function of F

(ii.) For α ∈ (0, 1), the α-quantile of F is given by

qα (F ) := F ← (α) := inf{x ∈ R : F (x) ≥ α}

Remarks
if X is a rv with df F we set xα := qα (F )
if F is continuous and strictly increasing xα = qα (F ) = F −1 (α),
where F −1 is the ordinary inverse of F

Risk Measures based on loss distribution March 26, 2022 10

 Definition of VaR
Value-at-Risk

Definition (VaR). Given some confidence level α ∈ (0, 1), Value-at-Risk

(VaR) of a portfolio with loss L at level α is defined as

VaRα = VaRα (L) = FL← (α) := inf{` ∈ R : FL (`) ≥ α} (1)

that is, VaR is the smallest number ` such that P(L > `) ≤ 1 − α.

VaR is simply the α-quantile of the loss distribution (typically, we compute

VaR0.95 , VaR0.99 )
Market-risk: ∆t = 10 days; Credit/operational: ∆t=1 year
VaR gives no information about the severity of losses occurring with a
probability less than 1 − α
VaR needs to be estimated from data

Risk Measures based on loss distribution March 26, 2022 11

 Mean-VaR
Value-at-Risk

Let µ be the mean of the loss distribution. The statistic

VaRm
α := VaRα − µ

denotes the mean-VaR and is used for capital-adequacy purposes instead

of ordinary VaR.
if ∆t = 1 day, then VaRm
α is referred to as daily earnings at risk
Example: in loan pricing one uses VaRm α to determine the economic
capital needed as a buffer against unexpected losses in a loan
portfolio

Risk Measures based on loss distribution March 26, 2022 12

 Table of Contents
VaR: Examples

1 Risk Measures
Introduction
Value-at-Risk
VaR: Examples
Expected Shortfall
ES: Examples
ES vs VaR

2 VaR and ES: Properties

Non-subadditivity
Coherent risk measures

3 Other Risk Measures Based on Loss Distributions

Risk Measures based on loss distribution March 26, 2022 13

 VaR for exponential distributions
VaR: Examples

Consider an insurance loss random variable L with an exponential

distribution having mean θ > 0:
1
exp(−`/θ); FL (`) = 1 − e −`/θ , for ` > 0.
f (`) =
θ
Given α ∈ (0, 1), VaRα (L) must be the value `α satisfying

α = FL (`α ) = P(L ≤ `α ) = 1 − exp{−`α /θ}

Hence
VaRα (L) = FL−1 (α) = −θ log(1 − α).
Remark: the VaR of any continuous random variables is simply the
inverse of the corresponding cdf.

Risk Measures based on loss distribution March 26, 2022 14

 VaR for normal and lognormal distributions
VaR: Examples

Suppose L ∼ N(µ, σ 2 ). Then, for a fixed α ∈ (0, 1)

VaRα (L) = µ + Φ−1 (α) σ,

where Φ denotes the standard normal df and Φ−1 (α) is the α-quantile of
Φ. Clearly, P(L ≤ VaRα (L)) = α.

Risk Measures based on loss distribution March 26, 2022 15

 VaR for normal and lognormal distributions
VaR: Examples

Suppose L ∼ N(µ, σ 2 ). Then, for a fixed α ∈ (0, 1)

VaRα (L) = µ + Φ−1 (α) σ,

where Φ denotes the standard normal df and Φ−1 (α) is the α-quantile of
Φ. Clearly, P(L ≤ VaRα (L)) = α.

Remark: the VaR of a linear transformation is equivalent to the linear

transformation of the VaR:
If Z ∼ N(0, 1) → VaRα (Z ) = Φ−1 (α)
If X = µ + Z σ → VaRα (X ) = µ + VaRα (Z )σ
This is in general true as long as the transformation is strictly increasing

Risk Measures based on loss distribution March 26, 2022 15

 VaR for normal and lognormal distributions (cont)
VaR: Examples

Suppose L ∼ N(µ, σ 2 ). Let

g (L) = Y = exp(L). Then lognormal densities

µ = 1, σ = 1

0.30
µ = 1, σ = 0.5

Y ∼ LogNormal(µ, σ 2 ) µ = 1.5, σ = 0.5

0.25
0.20
i.e. Y has a lognormal distribution

density
with parameters µ ∈ R and σ 2

0.15
0.10
(σ > 0).

0.05
For α ∈ (0, 1), the VaR of Y = e L is
0.00
0 5 10 15

VaRα (Y ) = e VaRα (L) = exp(Φ−1 (α) σ+µ).

Risk Measures based on loss distribution March 26, 2022 16

 VaR for Student’s t-distributions
VaR: Examples

Suppose L∗ = (L − µ)/σ ∼ tν , that is L∗ has a “standard” Student’s

t−distribution with ν degrees of freedom (benchmark model in finance,
usually ν = 3, 5):

Γ((ν + 1)/2)
f (x) = √ (1 + x 2 /ν)−(ν+1)/2 , −∞ < x < ∞, ν > 0,
πν Γ(ν/2)
R∞
where Γ(u) = 0 t u−1 e −t dt. For L ∼ t(µ, µ, σ 2 ) we have E (L) = µ and
V (L) = νσ 2 /(ν − 2) when ν > 2 (σ is not the standard deviation of the
distribution. We get
VaRα = µ + σtν−1 (α),
where tν denotes the df of standard t with ν dof, and tν−1 is its inverse.

Risk Measures based on loss distribution March 26, 2022 17

 Illustration: VaR for Skew t-distributions
VaR: Examples

0.35
VaR0.95

0.30
0.25
0.20
Loss density

0.15
0.10
0.05
0.00

−4 −2 0 2 4 6 8

Loss

Figure: VaR (α = 0.95) for a skew t4 -distribution

Risk Measures based on loss distribution March 26, 2022 18

 Table of Contents
Expected Shortfall

1 Risk Measures
Introduction
Value-at-Risk
VaR: Examples
Expected Shortfall
ES: Examples
ES vs VaR

2 VaR and ES: Properties

Non-subadditivity
Coherent risk measures

3 Other Risk Measures Based on Loss Distributions

Risk Measures based on loss distribution March 26, 2022 19

 Definition of Expected Shortfall (ES)
Expected Shortfall

VaR does not reflect the extremal losses occurring beyond the
(1 − α) × 100% chance worst scenario. Expected shortfall was introduced
by Artzner et al. (1997) (see also, Artzner et al. (1999))

Definition (ES). For a loss L with E (|L|) < ∞ and df FL , the ES at

confidence level α ∈ (0, 1) is defined as
Z 1 Z 1
1 1
ESα = qu (FL )du = VaRu (L)du (2)
1−α α 1−α α

where qu (FL ) = FL← (u) is the quantile function of FL .

ES is also known as conditional value at risk (CVaR)

Risk Measures based on loss distribution March 26, 2022 20

 Remarks on the ES measure
Expected Shortfall

ES is obtained by averaging VaR, for all u ≥ α (average loss when VaR is

exceeded), hence ES depends on FL and

ESα (L) ≥ VaRα (L)

If FL is continuous, then VaR can be viewed as the expected loss

that is incurred in the event that VaR is exceeded
Z ∞
1
ESα (L) = E(L|L > VaRα (L)) = `f (`)d`
1 − α qα (L)

ESα gives information about frequency and size of large losses (that
occur when the VaR “bad times” threshold has been exceeded)

Risk Measures based on loss distribution March 26, 2022 21

 Table of Contents
ES: Examples

1 Risk Measures
Introduction
Value-at-Risk
VaR: Examples
Expected Shortfall
ES: Examples
ES vs VaR

2 VaR and ES: Properties

Non-subadditivity
Coherent risk measures

3 Other Risk Measures Based on Loss Distributions

Risk Measures based on loss distribution March 26, 2022 22

 ES for exponential distributions
ES: Examples

Assume L ∼ Exp(1/θ), so that E (L) = θ.

For α ∈ (0, 1), we found VaRα (L) = −θ log(1 − α). Hence, we obtain

Z 1
1
ESα = VaRu (L)du
1−α α
Z 1
θ
=− log(1 − u)du
1−α α
Z 1−α
θ
=− log(y )dy
1−α 0
= −θ log(1 − α) + θ = VaRα + θ

Risk Measures based on loss distribution March 26, 2022 23

 ES for normal loss distributions
ES: Examples

Suppose L ∼ N(µ, σ 2 ). Then, for a fixed α ∈ (0, 1)

 !
L−µ L−µ L−µ
ESα = µ + σE ≥ qα = µ + σESα (L∗ ),
σ σ σ

L−µ
where L∗ := . We get
σ Z ∞ Z ∞
∗ 1 2
(1 − α)ESα (L ) = `φ(`)d` = ` √ e −` /2 d`
Φ−1 (α)
Z ∞ Φ−1 (α) 2π
1
=√ e −x dx, ω(α) = (Φ−1 (α))2 /2
2π ω(α)
1
= √ e −ω(α) = φ(Φ−1 (α)).
2π
−1
Hence, ESα = µ + σ φ(Φ1−α(α))
Risk Measures based on loss distribution March 26, 2022 24
 ES for Student’s t-distributions
ES: Examples

Suppose L∗ = (L − µ)/σ ∼ tν , that is L∗ has a “standard” Student’s

t−distribution with ν degrees of freedom (ν > 1)
tν , fν are the cdf and the density of standard t, respectively
t −1 := tν−1 (α)
ESα (L) = µ + σESα (L∗ )
fν (t −1 ) ν + (t −1 )2
 
ESα (L∗ ) =
1−α ν−1
Therefore
fν (t −1 ) ν + (t −1 )2
 
ESα (L) = µ + σ
1−α ν−1

Risk Measures based on loss distribution March 26, 2022 25

 Exercise
ES: Examples

ES of a lognormal distribution
Consider an insurance loss random variable L ∼ logN (µ, σ 2 ). Show that
2
e µ+σ /2
ESα (L) = Φ(Φ−1 (α) − σ)
1−α
where Φ(·) is the cdf of a standard normal rv.

Risk Measures based on loss distribution March 26, 2022 26

 Table of Contents
ES vs VaR

1 Risk Measures
Introduction
Value-at-Risk
VaR: Examples
Expected Shortfall
ES: Examples
ES vs VaR

2 VaR and ES: Properties

Non-subadditivity
Coherent risk measures

3 Other Risk Measures Based on Loss Distributions

Risk Measures based on loss distribution March 26, 2022 27

 Shortfall-to-quantile ratio
ES vs VaR

The difference between VaR and ES matters for a heavy-tailed

distribution:
Normal limα→1 ESα = 1
VaRα

Student-t limα→1 ESα = ν−1 ν

>1
VaRα
(with ν = 3, ES is 50% larger than VaR in the limit for
large α).

→ ESα is sensitive to the severity of losses exceeding VaRα .

Risk Measures based on loss distribution March 26, 2022 28

 VaR and ES for stock returns
ES vs VaR

Suppose the current value of a position on a particular stock is

Vt = 10000.
Assume Xt+1 represents daily log-returns on the stock. The (linearized)
loss for this portfolio is
L∆t+1 = −Vt Xt+1

We assume that
Xt+1 has zero mean
√
Xt+1 standard deviation σX = 0.2/ 250 (annualized volatility of
20%)
We compare VaR and ES under
1. a normal distribution µ = 0, σ = Vt σX
2. a t-distribution with ν = 5 dof scaled to have standard deviation σX

Risk Measures based on loss distribution March 26, 2022 29

 VaR and ES for stock returns
ES vs VaR
VaR and ES for α = 0.9, 0.99, under the normal model (risk measures
computed via the R package qrmtools [Hofert et al. (2021)])

VaR0.9
0.0030
ES0.9
0.0025
0.0020
Loss density (normal)

0.0015
0.0010
0.0005
0.0000

−300 −200 −100 0 100 200 300 400

Loss

Risk Measures based on loss distribution March 26, 2022 30

 VaR and ES for stock returns
ES vs VaR
VaR and ES for α = 0.9, 0.99 under the normal model (risk measures
computed via the R package qrmtools [Hofert et al. (2021)])

VaR0.9
0.0030
ES0.9
0.0025
0.0020
Loss density (normal)

0.0015
0.0010
0.0005

VaR0.99
0.0000

ES0.99

−300 −200 −100 0 100 200 300 400

Loss

Risk Measures based on loss distribution March 26, 2022 31

 VaR and ES for stock returns
ES vs VaR

The t distribution is a symmetric distribution with heavy tails

→ large absolute values are much more probable than in the normal
model.

Is the t model riskier than the normal model?

Risk Measures based on loss distribution March 26, 2022 32

 VaR and ES for stock returns
ES vs VaR

The t distribution is a symmetric distribution with heavy tails

→ large absolute values are much more probable than in the normal
model.

Is the t model riskier than the normal model?

α 0.9 0.95 0.99 0.995

VaR (normal) 162.10 208.10 294.30 325.80
VaR (t) 144.60 197.40 329.70 395.10
ES (normal) 222.00 260.90 337.10 365.80
ES (t) 225.60 283.20 436.20 514.40

Risk Measures based on loss distribution March 26, 2022 32

 VaR and ES for stock returns: remarks
ES vs VaR

The normal distribution appears to be at least as risky as the t model

using VaR at the 95% or 97.5% confidence level
::::

Risk Measures based on loss distribution March 26, 2022 33

 VaR and ES for stock returns: remarks
ES vs VaR

The normal distribution appears to be at least as risky as the t model

using VaR at the 95% or 97.5% confidence level
::::
Only for higher levels (e.g. 99%) the higher risk in the tails of the t
model become apparent
::::

Risk Measures based on loss distribution March 26, 2022 33

 VaR and ES for stock returns: remarks
ES vs VaR

The normal distribution appears to be at least as risky as the t model

using VaR at the 95% or 97.5% confidence level
::::
Only for higher levels (e.g. 99%) the higher risk in the tails of the t
model become apparent
::::
Expected shortfall better reflects the risk in the tails of the t model for
lower values of α

Risk Measures based on loss distribution March 26, 2022 33

 VaR and ES for stock returns
ES vs VaR

Comparison between ES/VaR under the normal and t model (the smaller
the degrees of freedom the heavier the tails...)
1000

600
400
500

200
0
0

−200
−400
−500

ESα for t 3 model ESα for t 5 model

VaRα for t 3 model VaRα for t 5 model
ESα for normal model ESα for normal model
−600

VaRα for normal model VaRα for normal model

0.001 0.005 0.050 0.500 0.001 0.005 0.050 0.500

1−α 1−α

Risk Measures based on loss distribution March 26, 2022 34

 Table of Contents

1 Risk Measures
Introduction
Value-at-Risk
VaR: Examples
Expected Shortfall
ES: Examples
ES vs VaR

2 VaR and ES: Properties

Non-subadditivity
Coherent risk measures

3 Other Risk Measures Based on Loss Distributions

Risk Measures based on loss distribution March 26, 2022 35

 Table of Contents
Non-subadditivity

1 Risk Measures
Introduction
Value-at-Risk
VaR: Examples
Expected Shortfall
ES: Examples
ES vs VaR

2 VaR and ES: Properties

Non-subadditivity
Coherent risk measures

3 Other Risk Measures Based on Loss Distributions

Risk Measures based on loss distribution March 26, 2022 36

 VaR pros and cons
Non-subadditivity

Some advantages of utilizing VaR include

possessing a clear interpretation and a relatively simple computation
for many distributions with closed-form df
obtaining VaR of strictly increasing functions by means of the same
transformation on the VaR of the original rv
no additional assumption required

Risk Measures based on loss distribution March 26, 2022 37

 VaR pros and cons
Non-subadditivity

Some advantages of utilizing VaR include

possessing a clear interpretation and a relatively simple computation
for many distributions with closed-form df
obtaining VaR of strictly increasing functions by means of the same
transformation on the VaR of the original rv
no additional assumption required
Major limitations of VaR are
we can lose (much) more than VaR, depending on the heaviness of
the tail of the loss distribution
VaR is not a coherent risk measure implying that diversification
benefits may not be fully reflected

Risk Measures based on loss distribution March 26, 2022 37

 Non-subadditivity of VaR
Non-subadditivity

Consider two loss distributions FL1 and FL2 for two portfolios; the overall
loss distribution of the merged portfolio L = L1 + L2 is FL . It is not
guaranteed that
qα (FL ) ≤ qα (FL1 ) + qα (FL2 )
Hence the VaR of the merged portfolio is not necessarily bounded above
by the sum of the VaRs of the individual portfolios.
This implies that
a diversification benefit associated with merging the portfolios is not
reflected by VaR
we cannot be sure that by aggregating VaR numbers for different
portfolios we will obtain a bound for the overall risk of the enterprise.

Risk Measures based on loss distribution March 26, 2022 38

 Non-subadditivity: example
Non-subadditivity

Let L1 , L2 ∼ Exp(1) and L1 , L2 independent

(P(L1 > x) = P(L2 > x) = exp(−x)). Then, it can be shown that VaR
is superadditive, that is,

VaRα (L1 + L2 ) > VaRα (L1 ) + VaRα (L2 )

for α < 0.71, and subadditive

VaRα (L1 + L2 ) ≤ VaRα (L1 ) + VaRα (L2 )

otherwise.

Risk Measures based on loss distribution March 26, 2022 39

 Illustration: non-subadditivity of VaR
Non-subadditivity

Independent L1, L2 ~ Exp(1)

VaRα(L1 + L2)
VaRα(L1) + VaRα(L2)

20.00
5.00
2.00
VaRα

0.50
0.20
0.05

Superadditivity Subadditivity
0.02

0.0 0.2 0.4 0.6 0.8 1.0

Risk Measures based on loss distribution March 26, 2022 40

 More on Non-subadditivity of VaR
Non-subadditivity

Formally, consider a risk measure

ψ : L → ψ(L)

for loss L ∈ M, a linear space of random variables, which include

constants.
Subadditivity ∀ L1 , L2 ∈ M : ψ(L1 + L2 ) ≤ ψ(L1 ) + ψ(L2 )
::::
Does it matter if a risk measures is subadditive or not?
Subadditivity reflects the idea that risk can be reduced by diversification,
and makes decentralization of risk-management systems possible

Risk Measures based on loss distribution March 26, 2022 41

 More on Non-subadditivity of VaR (cont)
Non-subadditivity

Example Consider two trading desks with positions leading to losses L1

and L2 .
Suppose a risk manager wants to ensure that the risk of the overall loss

L = L1 + L2

is smaller than some number q: if the chosen ψ is subadditive, then find

q1 , q2 such that
ψ(L1 ) ≤ q1 , ψ(L2 ) ≤ q2
and q1 + q2 ≤ q. Hence

ψ(L) = ψ(L1 + L2 ) ≤ q1 + q2 ≤ q.

Risk Measures based on loss distribution March 26, 2022 42

 VaR for a portfolio of defaultable bonds
Non-subadditivity

Example: Consider a portfolio of d = 100 defaultable corporate bonds:

assume that defaults of different bonds are independent and default
probability is p = 2%.
V0 = 100 current price of the bonds
If there is no default, a bond pays in t + 1 (one year) an amount of
105
Let Li be the loss of bond i, then

Li = 100Yi − 5(1 − Yi ) = 105Yi − 5

where Yi = 1 if bond i defaults in [t, t + 1], and 0 otherwise.

Risk Measures based on loss distribution March 26, 2022 43

 VaR for a portfolio of defaultable bonds (cont)
Non-subadditivity

The losses Li form a sequence of iid rvs

P(Li = −5) = P(Yi = 0) = 1 − p = 0.98

P(Li = 100) = P(Yi = 1) = p = 0.02

We compare two portfolios, both with current value equal to 10000:

Portfolio A 100 units of bond one (fully concentrated)

LA = 100L1

Portfolio B one unit of each of the bonds (diversified)

100
X
LB = Li
i=1

Risk Measures based on loss distribution March 26, 2022 44

 VaR for a portfolio of defaultable bonds (cont)
Non-subadditivity

We want to compute VaR at a confidence level of 95% for both

portfolios:
100
!
X
VaR0.95 (LA ) = 100VaR0.95 (L1 ); VaR0.95 (LB ) = VaR0.95 105Yi − 5
i=1

and VaR0.95 (L1 ) = −5. Moreover,

VaR0.95 (LB ) = 105 q0.95 (S) − 500

where S = 100
P
i=1 Yi ∼ Bin(100, 0.02). Since, P(S ≤ 5) ≈ 0.985 and
P(S ≤ 4) ≈ 0.949 < 0.95, q0.95 (S) = 5. Hence we get

VaR0.95 (LA ) = 100(−5) = −500, VaR0.95 (LB ) = 525 − 500 = 25

Risk Measures based on loss distribution March 26, 2022 45

 VaR for a portfolio of defaultable bonds (cont)
Non-subadditivity

VaR is not subadditive

100 100
!
X X
25 = VaR0.95 Li > VaR0.95 (Li ) = −500
i=1 i=1

The risk capital required for portfolio B is higher than for portfolio A.
Moreover
even after a withdrawal of a risk capital of 500, portfolio A is still
acceptable to a risk controller working with VaR at the 95% level
an additional risk capital of 25 is required for portfolio B to satisfy a
regulator working with VaR at the 95% level
This contradicts the fact that portfolio B should have a lower VaR being
less risky than portfolio A.

Risk Measures based on loss distribution March 26, 2022 46

 VaR for a portfolio of defaultable bonds: Remarks
Non-subadditivity

In the last example, the non-subbaditivity of VaR is caused by the fact

that the assets making up the portfolio have very skewed loss
distributions. Non-subadditivity of VaR also occurs
when the underlying rvs are independent but very heavy-tailed
for dependent losses, when their dependence structure is highly
asymmetric
However, it can be shown that VaR is subadditive in the situation where
all portfolios can be represented as linear combinations of the same set of
underlying elliptically distributed (e.g., normally distributed) risk factors

Risk Measures based on loss distribution March 26, 2022 47

 Table of Contents
Coherent risk measures

1 Risk Measures
Introduction
Value-at-Risk
VaR: Examples
Expected Shortfall
ES: Examples
ES vs VaR

2 VaR and ES: Properties

Non-subadditivity
Coherent risk measures

3 Other Risk Measures Based on Loss Distributions

Risk Measures based on loss distribution March 26, 2022 48

 Coherent risk measure
Coherent risk measures

Axioms for coherent risk measures

(A1) Monotonicity ∀ L1 , L2 ∈ M, L1 ≤ L2 almost surely: ψ(L1 ) ≤ ψ(L2 )
(A2) Translation invariance ∀ L ∈ M, ` ∈ R: ψ(L + `) = ψ(L) + `
(A3) Positive homogeneity ∀ L ∈ M, λ > 0: ψ(λL) = λψ(L)
(A4) Subadditivity ∀ L1 , L2 ∈ M : ψ(L1 + L2 ) ≤ ψ(L1 ) + ψ(L2 )

Risk Measures based on loss distribution March 26, 2022 49

 Coherent risk measure
Coherent risk measures

Axioms for coherent risk measures

(A1) Monotonicity ∀ L1 , L2 ∈ M, L1 ≤ L2 almost surely: ψ(L1 ) ≤ ψ(L2 )
(A2) Translation invariance ∀ L ∈ M, ` ∈ R: ψ(L + `) = ψ(L) + `
(A3) Positive homogeneity ∀ L ∈ M, λ > 0: ψ(λL) = λψ(L)
(A4) Subadditivity ∀ L1 , L2 ∈ M : ψ(L1 + L2 ) ≤ ψ(L1 ) + ψ(L2 )

Remark
I VaR always satisfies (A1)–(A3), but not (A4), in general.
I Expected shortfall is a coherent risk measure (proof omitted)

Risk Measures based on loss distribution March 26, 2022 49

 VaR is coherent in the normal case
Coherent risk measures

The computation of VaR(L1 + L2 ) requires assumptions on the marginals

and the dependence between the risks (joint distribution).
If (L1 , L2 ) ∼ N2 (µ, Σ), then Li ∼ N (µi , σi2 ), i = 1, 2, where
 2 
σ1 ρσ1 σ2
µ = (µ1 , µ2 ), Σ = , ρ ∈ [−1, 1] (3)
ρσ1 σ2 σ22
and
L1 + L2 ∼ N (µ1 + µ2 , σ12 + σ22 + 2ρσ1 σ2 )
Let α > 0.5, then
q
VaRα (L1 + L2 ) = µ1 + µ2 + σ12 + σ22 + 2ρσ1 σ2 Φ−1 (α)
q
≤ µ1 + µ2 + (σ1 + σ2 )2 Φ−1 (α)
= (µ1 + σ1 Φ−1 (α)) + (µ2 + σ2 Φ−1 (α))
= VaRα (L1 ) + VaRα (L2 )

Risk Measures based on loss distribution March 26, 2022 50

 Further comments: Choice of VaR parameters
Coherent risk measures

There are two important choices when working with VaR:

Choice of ∆t should reflect the time period over which a financial
institution is committed to hold its portfolio
usually one year for measuring the risk in the liability and
asset portfolios of an insurer
∆t should be relatively small to (i) use of the linearized loss
operator (ii) assume the composition of the portfolio
remains unchanged

Risk Measures based on loss distribution March 26, 2022 51

 Further comments: Choice of VaR parameters
Coherent risk measures

There are two important choices when working with VaR:

Choice of ∆t should reflect the time period over which a financial
institution is committed to hold its portfolio
usually one year for measuring the risk in the liability and
asset portfolios of an insurer
∆t should be relatively small to (i) use of the linearized loss
operator (ii) assume the composition of the portfolio
remains unchanged

Choice of α can be different according to the specific purpose: for

instance, the Basel Committee proposes the use of VaR at
the 99% level and ∆t = 10 days for market risk
in general, capital-adequacy purposes require a high
confidence level in order to have a sufficient safety margin

Risk Measures based on loss distribution March 26, 2022 51

 Table of Contents

1 Risk Measures
Introduction
Value-at-Risk
VaR: Examples
Expected Shortfall
ES: Examples
ES vs VaR

2 VaR and ES: Properties

Non-subadditivity
Coherent risk measures

3 Other Risk Measures Based on Loss Distributions

Risk Measures based on loss distribution March 26, 2022 52

 Variance as a risk measure

The variance of the P&L distribution has been used as a risk measure in
finance:
is a well-understood concept which is easy to use analytically
in market risk, the volatility can be interpreted as a measure of
uncertainty
However,
we have to assume that the second moment of the loss distribution
exists
it makes no distinction between positive and negative deviations
from the mean
variance is a good measure of risk only for distributions which are
(approximately) symmetric around the mean, such as the normal
distribution or a (finite-variance) Student’s t-distribution

Risk Measures based on loss distribution March 26, 2022 53

 Alternative risk measures

Some risk measures have been proposed that are simultaneously coherent
and may also consider losses beyond VaR.
In Acerbi and Tasche (2002) are listed five measures of risk that include
losses in excess of VaR
Conditional VaR (CVaR)
Expected shortfall (ES)
Tail conditional expectation (TCE)
Worst conditional expectation (WCE)
Spectral risk measures

Risk Measures based on loss distribution March 26, 2022 54

 First R lab

We use R compute VaR and ES for the different methods:

Variance-covariance method (assume that the linearized loss provides
a sufficiently accurate approximation and multivariate normal)
Historical simulation method (using empirically estimated risk
measures)
Monte Carlo simulation method (simulate losses from fitted
multivariate t or normal risk-factor changes)

Risk Measures based on loss distribution March 26, 2022 55

 Acerbi, C. and Tasche, D. (2002). On the coherence of expected shortfall. Journal of
Banking & Finance, 26:1487–1503.
Artzner, P., Delbaen, F., Eber, J. M., and Heath, D. (1997). Thinking coherently.
Risk, 10(11):68–71.
Artzner, P., Delbaen, F., Eber, J. M., and Heath, D. (1999). Coherent measures of
risk. Mathematical Finance, 9:203–228.
Hofert, M., Hornik, K., and McNeil, A. J. (2021). qrmtools: Tools for Quantitative
Risk Management. R package version 0.0-14.
Jorion, P. (2007). Value at Risk: The New Benchmark for Managing Financial Risk.
McGraw-Hill, New York, 3 edition.

Risk Measures based on loss distribution March 26, 2022 56