Non Life Insurance Premium Rating

José Garrido
Department of Mathematics and Statistics
Concordia University, Montreal, Canada

Fall 2020
c
José Garrido, 2020
2
Chapter 1

Risk measures

Risk measures are defined and used on gains, for example on investments,
and on losses, for example on investment or insurance losses. In this course
we concentrate on measures used to asses the potential risk for insurance
losses.
Let X : Ω → R be the loss variable (risk) associated with a given
insurance, over a single period of time 0 to T . Then let X be the set of all
risks. A risk measure is defined as follows.
Definition 1.1 A risk measure is a function ρ : X → R.
In insurance, ρ(X) measures the risk in terms of potential losses; that is
an amount of Euros that can be lost over a period of reference.

1.1 Value at risk (V aR)

One well known risk measure is Value at Risk (V aR1−p). It can be defined
as:
For a given time horizon T and an (1 − p) × 100% confidence
level, V aR1−p is the loss value that can only be exceeded with a
probability of at most p∗ .
V aR1−p usually answers the question: What is the minimal loss incurred
in the p × 100% worse outcomes of the insurance portfolio?
Mathematically, V aR1−p is simply the (1 − p) × 100%–percentile of the
loss distribution. Usually p is chosen to be a small value, p = 1% or p = 5%.
∗
Duffie and Pan (1997).

3
4 CHAPTER 1. RISK MEASURES

Figure 1.1: VaR

Source: https://analystprep.com/study-notes/frm/part-1/
valuation-and-risk-management/measures-of-financial-risk/

Definition 1.2 For a possible loss X over a given period [0, T ] and 0 < p <
1, the (1 − p) × 100% Value at Risk is:
V aR1−p(X) = sup{x ∈ R | P(X ≥ x) > p}.
Example 1.1 Pareto(α, θ):
Let us derive V aR0.95(X) for a loss distribution commonly used in insurance
for heavy–tailed risks, X ∼ Pareto(α, θ). Here say α = 2 and θ = 10, 000:
 α  2
θ 10, 000
P(X ≤ x) = FX (x) = 1 − = 1− ,
θ+x 10, 000 + x
so
V aR0.95(X) = sup{x ∈ R | P(X ≥ x) > 0.05}
= sup{x ∈ R | 1 − FX (x) > 0.05}
= sup{x ∈ R | FX (x) ≤ 0.95},
that is equivalent to finding the inverse y = V aR0.95(X) = FX−1(0.95), that is
 2
10, 000
1− = 0.95.
10, 000 + y
1.1. VALUE AT RISK (V AR) 5
√
Isolating gives y = V aR0.95(X) = (10, 000/ 0.05) − 10, 000 = 34, 721.36.

Remark 1.1 Note that the relation V aR1−p (X) = FX−1(1 − p) obtained here
for the Pareto actually holds for any continuous loss distribution.

Exercise 1.1 Redo Example 1.1 (a) if the loss X has an exponential (light–
tailed) distribution with mean 10,000 and (b) when X is normal (also light–
tailed), with mean 10,000 and variance 100,000,000 (you can use Excel here
to calculate V aR0.95(X)).

One main advantage of V aR1−p is that, irrespectively of the insurance

risk to which it is applied, it is always expressed as money lost. In addi-
tion V aR1−p is simple to calculate for a wide variety of risks, explaining why
regulators opted for it in setting solvency standards. But despite its univer-
sality, several authors have pointed out the deficiencies of V aR1−p: it is not
sub–additive, nor convex and difficult to optimize.
Artzner et al. (1997) consider how risk measures should behave. They
suggest a set of properties that coherent risk measure should satisfy.

Definition 1.3 [Artzner et al. (1997)]

A risk measure ρ is called coherent if and only if it satisfies the following
axioms:

1. Sub–additivity: for any X, Y ∈ X, then ρ(X + Y ) ≤ ρ(X) + ρ(Y ),

2. Positive homogeneity: for any X ∈ X and λ ≥ 0, then ρ(λ X) =

λ ρ(X).

3. Translation invariance: for a fixed X ∈ X and any a ∈ R, then ρ(X +

a) = ρ(X) + a∗

4. Monotonicity: let X, Y ∈ X be such that X ≤ Y a.s., then ρ(X) ≤

ρ(Y ).

∗
The property would be ρ(X + a) = ρ(X) − a for assets rather than losses.
6 CHAPTER 1. RISK MEASURES

1.2 Conditional value at risk (CV aR)

An alternative to V aR1−p is the so–called Conditional Value at Risk (CV aR1−p),
also related to Conditional Tail Expectation (CT E), Average Value at Risk
(AV aR) or Expected Short–fall (ES), although these can be misnomers in
some cases. It was initially proposed by Artzner et al. (1999) and has been
studied extensively since.
CV aR1−p answers the question: What is the expected loss incurred in
the p × 100% worse cases in closing a position?
Definition 1.4 [Artzner et al. (1999)]
For a any risk X and a confidence level 0 < 1 − p < 1, the (1 − p) × 100%
CV aR1−p is given by:
1 p
Z
CV aR1−p (X) = V aR1−t(X)dt.
p 0
By definition, CV aR1−p (X) ≥ V aR1−p(X), for any loss risk X. In fact,
for continuously distributed losses it is characterized as
 
CV aR1−p (X) = E X | X > V aR1−p (X) .
Arguably, the properties of CV aR1−p are more attractive than those of
V aR1−p: it is sub–additive and convex. In fact, it is in most cases coherent
[Artzner et al. (1999)]. As we will see below, it satisfies additional interesting
properties.
Example 1.2 Exponential(θ) revisited:
Reconsider the case in Exercise 1.1(a) where X ∼ exponential(θ = 10, 000)
and p = 0.05. Then from the definition above
Z 0.05
1
CV aR0.95(X) = V aR1−t (X)dt,
0.05 0
where, by Remark 1.1, for the exponential distribution we have V aR1−t(X) =
FX−1 (1 − t) = −θ ln(t) = −10, 000 ln(t). So
Z 0.05
CV aR0.95(X) = 20(−10, 000) ln (t)dt
0
 0.05
= −200, 000 t ln(t) − t 0
 
= −200, 000 (0.05) ln(0.05) − 0.05
= −5, 000[ln(0.05) − 1] = 190, 000(0.0513) = 19, 978.66.
1.2. CONDITIONAL VALUE AT RISK (CV AR) 7

Exercise 1.2 (a) Calculate CV aR0.95(X) for the normal (µ = 10, 000, σ 2 =
100, 000, 000) example above and compare the value you obtain to the
V aR0.95(X) value you derived in Exercise 1.1(b). Which one is higher?
Would your answer change if you compared V aR0.99(X) to the above
CV aR0.95(X).

(b) Similarly, derive CV aR0.95(X) for the Pareto loss X in Example 1.1
and compare to its V aR0.95(X) value. Again, which one is higher?

Other properties commonly applied to risk measures are:

5. Law invariance: ρ(X1 ) = ρ(X2 ) if X1 and X2 have the same distribu-

tion.

6. Additivity for comonotone risks: ρ(X + Y ) = ρ(X) + ρ(Y ) if X and Y

are comonotonic.†

7. Expectation bounded: ρ(X) ≥ E(X).

It can be seen that apart from satisfying the 4 properties of coherent risk
measures, CV aR1−p also satisfies Properties 5, 6 and 7 above.
One other measure based on CV aR is the Weighted CV aR (WCVaR)
defined as:
m
X
W CV aR1−p1 ,...,1−pm (X) = ki CV aR1−pi (X),
i=1

P
where 0 < pi < 1, and the weights ki > 0 are such that i ki = 1. Like
CV aR, the W CV aR risk measure also satisfies all the above properties, so in
particular, it is coherent. Its main advantage over CV aR is that by involving
several confidence levels 1 − pi , the measure uses additional parts of the loss
distribution, not just one percentile.
†
(X1 , X2 ) are said to be comonotonic if P{X1 ≤ x1 , X2 ≤ x2 } = mini=1,2 P{Xi ≤ xi }.
See Yaari (1987) for details.
8 CHAPTER 1. RISK MEASURES

1.3 Distortion measures

Pushing this idea of Weighted CV aR further, it turns out that there is a
better way to use fully all the information in the loss distribution. Wang
(1996) proposed to use risk measures obtained through distortion func-
tions. The following definitions make more precise Wang’s idea as well as
the smoothness and other restrictions for a function to be a valid distortion.

Definition 1.5 [Wang (1996)]

A distortion function is a continuous, strictly increasing and concave function
g : [0, 1] → [0, 1], such that g(0) = 0, g(1) = 1 and it is differentiable on [0, 1].

Wang (1996) defines a class of risk measures obtained through such distortion
functions. We give here an alternative but equivalent formulation.

Definition 1.6 Let g be a distortion function and X a loss. The correspond-

ing distortion risk measure is given by
Z 1
ρg (X) = V aR1−t (X)g 0 (t)dt.
0

All distortion risk measures ρg satisfy Properties 1 to 7, above. Hence, they

are all coherent. Special cases worth mentioning are the Power Transform
or Dual Power Transform (DPT ) with parameter a > 1, obtained with g(t) =
1 − (1 − t)a, and even more importantly, Wang’s risk measure (Wα ) obtained
with g(t) = Φ(Φ−1 (t) + a) for a > 0, where Φ is the normal cumulative
distribution function Z x
1 y2
Φ(x) = √ e− 2 dy
−∞ 2π
and Φ−1 : [0, 1] → (−∞, ∞) is its inverse. In both cases a can be seen as a
risk aversion parameter, that is, the larger a, the larger the risk aversion.

1.4 Spectral measures

Another interesting family of risk measures is generated, in part, by replacing
V aR by CV aR in the definition of distortion risk measures above.
1.5. DEVIATION MEASURES 9

Definition 1.7 [Acerbi (2002)] Pm

Let 0 < pi < 1, ki > 0 such that i=1 ki ≤ 1 and g : [0, 1] → [0, 1] be
a continuous function such that g(0) = 0, it is Pdifferentiable on [0, 1] and
non decreasing (that is g 0 (t) ≥ 0), with g(1) + m i=1 k1 = 1. The associated
spectral measure is given by
m
X Z 1
ρ1−p,k,g (X) = k1 CV aR1−pi (X) + CV aR1−t(X)g 0 (t)dt.
i=1 0

Spectral measures have been shown to satisfy Properties 1 to 7, without the

need to impose that g be concave or strictly increasing, which is significantly
less restrictive than for distortion risk measures. Some notable special cases
when m = 1 are the Spectral Dual Power Transform (SDP Ta ), for g(t) =
1 − (1 − t)a , and a > 0, as well as the Spectral Wang measure (SWa ), for
g(t) = Φ(Φ−1 (t) + a).

1.5 Deviation measures

Let σ(X) be a deviation of a loss X, for example the standard deviation,
absolute deviation, standard semi–deviation or absolute semi–deviation.

Definition 1.8 [Rockafellar et al. (2006)]

If σ(X) is a deviation on loss risk X then

ρσ (X) = σ(X) − E(X),

is its associated deviation risk measure.

It can be shown that deviation measures satisfy Properties 1, 2, 3, 5 and 7.

1.6 Entropic measures

 
Although analytically related to the classical entropy, E X ln(X) , of a ran-
dom variable X ≥ 0 (with the convention
 that
 0 ln(0)Y = 0), the entropic risk
measure is different. (Note that E X ln(X) = E(Y e ) for Y = ln(X)).
10 CHAPTER 1. RISK MEASURES

Definition 1.9 [Föllmer and Schied (2004)]

For loss risk X and parameter µ > 0, the entropic risk measure is given by

1  −µX 
ρµ (X) = ln E(e ) .
µ

It can be shown that the entropic measure satisfies Properties 3, 4, 5 and

7. Even if it is not positively homogeneous nor sub–additive (and hence not
coherent), ρµ (0) = 0 holds, which is a property close to sub–additivity plus
positive homogeneity, so close to convexity:

ρµ (tX1 + (1 − t)X2) ≤ tρµ (X1 ) + (1 − t)ρ(X2 ),

for 0 < t < 1. This “convexity”, together with a property of “temporal con-
sistency” (the study of which is beyond the scope of this course) makes this
entropic risk measure suitable for applications to dynamic problems. In fact
it is the only risk measure that satisfies temporal consistency, even though
other measures, such as CV aR, satisfy a “weak temporal consistency”.

Example 1.3 Entropic VaR:

Given that the entropic measure has the good temporal consistency property
but is not sub–additive, the problem remains open to find a sub–additive risk
measure “close” to the entropic measure. Consider the entropic V aR de-
fined, for parameters µ, c > 0, as
 
ρµ (tX) + c
EV aR(µ,c)(tX) = inf .
t>0 t

It can be shown that EV aR(µ,c)(tX) satisfies Properties 1, 2, 3, 4, 5 and 7,

hence it is coherent. From the above definition it is clear that the entropic
risk measure can be characterized through EV aR:

ρµ (X) = sup{EV aR(µ,c)(tX) − c}.

t>0

In other words, the entropic risk measure is the supremum of entropic V aRs
minus c, which means a certain similarity between the entropic V aR and the
entropic risk measure.
1.6. ENTROPIC MEASURES 11

Conclusions
Clearly, opinions differ substantially on the properties that risk measures
should satisfy. Desirable properties differ with the intended use for a risk
measure: capital requirements, statutory purposes, insurance or risk pre-
mium calculations. We can safely conclude that there is no unique general
set of axioms that is valid for all above applications. Certain uses will favor
some of the 7 properties listed here, while other applications call for other
properties.