Lecture Notes 2020
Lecture Notes 2020
Notes
1 Review of probability
1.1 Distribution. Tail distribution
Assume that X is a random variable with the cumulative distribution function (CDF) FX (x) =
P(X ≤ x) and the density fX (x) = F 0 (x). Then
Z x
F (x) = fX (u)du .
−∞
1
The distribution function of X is
0, x<0
FX (x) = P (X ≤ x) = .
1 − exp(−λx) , x ≥ 0
Pareto: We say that a random variable X is Pareto with parameter α > 0 if its density is given by
αx−α−1 , x ≥ 1 ,
fX (x) = . (2)
0, x<1.
Normal: We say that a random variable X is Normal with mean µ and variance σ 2 if its density is
given by
1
exp −(x − µ)2 /(2σ 2 ) , −∞ < x < ∞ .
fX (x) = √
2πσ
If µ = 0 and σ = 1 then we have the standard normal density denoted by φ(x). In this case the CDF
is denoted by Φ(x).
t (Student): We say that a random variable X has t distribution with α degrees of freedom if its
density is given by
− α+1
x2
Γ((α + 1)/2) 2
fX (x) = √ 1+ , −∞ < x < ∞ .
απΓ(α/2) α
Note that, similarly to Pareto distribution, we have
fX (x)
lim ∈ (0, ∞) .
x→∞ x−α−1
Thus, it is also an example of a heavy-tailed distribution.
2
1.3 Expected value
For a random variable X with a density fX the expected value is calculated as
Z ∞
E[X] = xfX (x) dx .
−∞
If X is a nonnegative random variable with the cumulative distribution function FX then there is an
alternative formula: Z ∞
E[X] = F̄ (x)dx .
0
Y = X + a: We calculate
FY (y) = P (Y ≤ y) = P (X + a ≤ y) = P (X ≤ y − a) = FX (y − a) .
fY (y) = fX (y − a) .
3
Y = exp(X) Since Y ≥ 0, we have P (Y ≤ y) = 0 for y ≤ 0. We calculate for y > 0:
Y = log(X): Note that we need that X is nonnegative (e.g. Pareto, exponential). We calculate
Note that starting with X Pareto, the transformation Y = log X leads to exponential random
variable.
4
Y = X 2: If X is nonnegative, then for x < 0, P (X 2 < x) = 0, while for x ≥ 0,
√ √
P (X 2 ≤ x) = P (X ≤ x) = FX ( x) .
The density of Y = X 2 is
1 √
fY (x) = √ fX ( x) .
2 x
√
Example 1.6 • Assume that X is Exponential. Then P (X 2 ≤ x) = 1 − exp(−λ x), x > 0. Thus,
for Y = X 2 ,
0. x<0
FY (x) = √ .
1 − exp(−λ x) , x ≥ 0
Y = X 2: If X is not nonnegative (like normal), then for x < 0, P (X 2 < x) = 0, while for x ≥ 0,
√ √ √ √
P (X 2 ≤ x) = P (− x < X ≤ x) = FX ( x) − FX (− x) .
The density of Y = X 2 is
1 √ √
fY (x) = √ fX ( x) + fX (− x) .
2 x
We note that the resulting random variable has no density, since there is a positive mass at some point.
This means that the cumulative distribution function has a jump at that point.
By the definition Y ≥ 0. Therefore, P (Y ≤ y) = 0 for all y < 0. For y = 0 we have
P (Y ≤ 0) = P (Y = 0) = P (X ≤ b) = FX (b) .
P (Y ≤ y) = P (Y ≤ y, X ≤ b) + P (Y ≤ y, X > b) = P (0 ≤ y, X ≤ b) + P (X − b ≤ y, X > b)
= P (X ≤ b) + P (b < X ≤ y + b) = P (X ≤ b) + P (X ≤ y + b) − P (X ≤ b)
= P (X ≤ y + b) = FX (y + b) . (9)
5
In summary
0, y<0
FY (y) = FX (b) , y=0 . (10)
FX (y + b) , y > 0
Note that the cumulative distribution function has the jump of size P (X ≤ b) at point y = 0.
• Assume that X is Pareto. We have to be careful about the value of b. Consider two cases: b < 1
and b > 1. See Assignment 1.
R∞
Expected value. We have the general formula for E[h(X)] = −∞ h(x)fX (x)dx. We use it with the
function h(x) = (x − b)+ which equals (x − b) if x > b and zero otherwise. We have
Z ∞ Z b Z ∞
E[(X − b)+ ] = (x − b)+ fX (x)dx = 0 × fX (x)dx + (x − b)fX (x)dx
−∞ −∞ b
Z ∞ Z ∞ Z ∞
= xfX (x)dx − b fX (x)dx = xfX (x)dx − bF̄X (b) .
b b b
In summary
Z ∞
E[(X − b)+ ] = xfX (x)dx − bF̄X (b) .
b
Thus
1
E[(X − b)+ ] = exp(−λb) .
λ
6
1.6 Inverse function. Quantile function.
Definition 1.9 Given a non-decreasing function T : R → R, the generalized inverse of T is T ← (y) :=
inf{x ∈ R : T (x) ≥ y}.
Definition 1.10 Given a distribution function F , the generalized inverse F ← is called the quantile
function of F .
If F is continuous and strictly increasing then the generalized inverse is just the usual inverse obtained
by solving y = F (x) for x. In this case we will write F −1 instead of F ← . Note that in this case
F −1 (F (y)) = F (F −1 (y)) = y .
In general
F ← (F (y)) 6= F (F ← (y)) 6= y .
Inverse function for transformations. Assume that X is random variable with the CDF FX . If
Y = h(X) and you want to find FY← then you proceed as follows:
1. Calculate FY in terms of FX ;
Y = log(X) Assume that X is a continuous and nonnegative random variable with the cumulative
distribution function FX that is strictly increasing. Then we know from (7) that
FY (x) = P (Y ≤ x) = FX (exp(x)) .
FX (exp(x)) = y ,
exp(x) = FX−1 (y) ,
x = log(FX−1 (y)) .
7
2 Risk measures
Definition 2.1 (Coherent Risk Measure) Let ρ be a risk measure. Let X, Y be random variables,
and let c ∈ R, t ≥ 0. We say that ρ is a coherent measure of risk if it satisfies the following four
properties.
2.1 Value-At-Risk
Definition 2.2 Let X be a loss variable with the distribution function FX . Given a probability p ∈
(0, 1), the Value-At-Risk of the loss variable X at the level p is given by the smallest number x such
that the probability that loss X exceeds x is no larger than p:
We will show that VaR satisfies properties 1-3 of Definition 2.1. In what follows we shall assume that
the corresponding cumulative distribution functions are continuous and strictly increasing.
Proof. Let X, Y be non-negative random variables with cumulative distribution functions FX and FY ,
respectively.
1. Let X ≤ Y . Then for x ∈ R we have FX (x) ≥ FY (x). This implies FX← (x) ≤ FY← (x). Setting
x = 1 − p we obtain the definition of VaR and therefore we conclude VaRX (p) ≤ VaRY (p) for
p ∈ (0, 1).
−1
2. We write VaRtX (p) = FtX (1 − p). We can also write the following relation:
(∗) (∗∗)
FtX (x) = FX (x/t) = 1 − p .
−1
From relation (*) we can conclude that x = FtX (1 − p). From relation (**) we can conclude that
− −
FX 1(1 − p) = x/t. This implies x = tFX 1(1 − p). Combining (*) and (**) together we get the
result:
VaRtX (p) = tVaRX (p) ,
as needed.
8
3. We write
(∗∗∗) (∗∗∗∗)
FX+c (y) = FX (y − c) = 1 − p .
−1
From relation (***) we can conclude that y = FX+c (1 − p) and from relation (****) we can
−1 −1
conclude that y − c = FX (1 − p). This implies y = FX (1 − p) + c. Combining these expressions
we get the result:
VaRX+c (p) = VaRX (p) + c ,
as needed.
However, as we will argue later, the Value-At-Risk is not a coherent risk measure, since in general
it is not sub-additive.
log(p)
VaRX (p) = − .
λ
Example 2.6 Let X be normal with mean µ and variance σ 2 . Let Φ be the standard normal distribu-
tion. Fix p ∈ (0, 1). We have
VaRX (p) = µ + σΦ−1 (1 − p) .
Lemma 2.8 Assume the CDF of X is continuous and strictly increasing. The expected shortfall can
be calculated as
1 1
Z
ESX (p) = VaRX (1 − u)du .
p 1−p
9
We have
E[XI{X > VaRX (p)}]
E[X|X > VaRX (p)] =
P (X > VaRX (p))
1 ∞
Z
1
= E[XI{X > VaRX (p)}] = xfX (x)dx
p p VaRX (p)
1 1 1 1
Z Z
−1
= F (u)du = VaRX (1 − u)du .
p 1−p X p 1−p
We used substitution x = FX−1 (u) and the property FX (FX−1 (x)) = x, valid since FX is strictly increasing
and continuous.
We can use both formulas for ESX , whatever is more convenient.
Example 2.9 Let X be Pareto random variable with parameter α. We have when α > 1
E[XI{X > y}] α
E[X|X > y] = = y.
F̄X (y) α−1
Hence, if α > 1, then ESX (p) is
α −1
ESX (p) = E[X|X > VaRX (p)] = p α .
α−1
Example 2.10 Let X be exponential random variable with parameter λ. Then
R∞ −λx dx
E[XI{X > y}] y λxe
E[X|X > y] = =
F̄ (y) e−λy
ye−λy + λ1 e−λy 1
= =y+ .
e−λy λ
Therefore,
1 − log(p)
ESX (p) = .
λ
Example 2.11 Let Z be standard normal. Then
φ(Φ−1 (1 − p))
ESZ (p) = .
p
Indeed, using a change of variables u = Φ(y) and du = φ(y)dy it is easy to calculate:
1 1 1 1 −1
Z Z
ESZ (p) = VaRZ (1 − u)du = Φ (u)du
p 1−p p 1−p
1 ∞
Z
= yφ(y)dy .
p Φ−1 (1−p)
You are asked to finish the computation in Assignment 2. Now, let X = µ+σZ. Then use the coherency
properties of ESX . See Assignment 2.
10
We will show for Expected Shortfall that it is indeed a coherent measure. Properties 1-3 will follow
directly from the results of Value-at-risk.
Lemma 2.12 The Expected Shortfall satisfies properties 1-4 of Definition 2.1.
1 1 1 1
Z Z
ESX (p) = VaRX (1 − u)du ≤ VaRY (1 − u)du = ESY (p) .
p 1−p p 1−p
2. See Assignment 2.
3. See Assignment 2.
1.
We write,
Define
M = (X − VaRX (p))(I{X ≥ VaRX (p)} − I{(X + Y ) ≥ VaRX+Y (p)}) .
We will show E[M ] ≥ 0. We have
(i) If X > VaRX (p) then we have I{X ≥ VaRX (p)} = 1 and by definition, I{(X + Y ) ≥ VaRX+Y (p)} ∈
[0, 1] which implies M ≥ 0.
(ii) If X < VaRX (p) then we have I{X ≥ VaRX (p)} = 0 and we have M ≥ 0.
(iii) If X = VaRX (p) then M = 0.
11
By (11) we can calculate
Similarly for Y ,
We conclude that
Since all four properties are satisfied we conclude that Expected Shortfall is indeed a coherent risk
measure.
12