Financial Risk: Credit Risk, Lecture 2

Alexander Herbertsson

Centre For Finance/Department of Economics

School of Economics, Business and Law, University of Gothenburg
E-mail: [email protected]
or
[email protected]

Financial Risk, Chalmers University of Technology,

Göteborg
Sweden

April 27, 2017

Alexander Herbertsson (Univ. of Gothenburg) Financial Risk: Credit Risk, Lecture 2 April 27, 2017 1 / 29
 Content of lecture

Short recapitulation of the mixed binomial model

Discussion of the loss distribution in the mixed binomial model and

how to use the LPA theory to find approximation for the loss for large
portfolios

Recapitulation of Value-at-Risk and Expected shortfall and its use in

the mixed binomial loss model

Study of a mixed binomial loss model with a beta distribution

Study of a mixed binomial loss model with a logit-normal distribution

Discussion of correlations etc.

Alexander Herbertsson (Univ. of Gothenburg) Financial Risk: Credit Risk, Lecture 2 April 27, 2017 2 / 29
 Recap of the mixed binomial model
Consider a homogeneous credit portfolio model with m obligors, and where each
obligor can default up to fixed time point, say T . Each obligor have identical
credit loss at a default, say ℓ. Here ℓ is a constant.
Let Xi be a random variable such that

1 if obligor i defaults before time T
Xi = (1)
0 otherwise, i.e. if obligor i survives up to time T
Let Z be a random variable, discrete or continuous, that represents some
common background variable affecting all obligors in the portfolio.
Since we consider a homogeneous credit portfolio, X1 , X2 , . . . Xm are
identically distributed. Furthermore, we assume the following:
Conditional on Z , the random variables X1 , X2 , . . . Xm are independent and
each Xi have default probability p(Z ) ∈ [0, 1], that is
P [ Xi = 1 | Z ] = p(Z ) (2)
so that P [Xi = 1] = p̄ for each obligor i where p̄ is given by
p̄ = E [Xi ] = E [E [ Xi | Z ]] = E [p(Z )] (3)

Alexander Herbertsson (Univ. of Gothenburg) Financial Risk: Credit Risk, Lecture 2 April 27, 2017 3 / 29
 Recap of the mixed binomial model, cont.

Note that (2) and (3) holds regardless if Z is a discrete or continuous

random variable.
If Z is a continuous random variable on R with density fZ (z) then
Z ∞
p̄ = E [p(Z )] = p(z)fZ (z)dz. (4)
−∞

Recall that we want to find the loss distribution in the homogeneous credit
portfolio specified on the previous slide.
The total credit loss in the portfolio at time T , called Lm , is
m
X m
X m
X
Lm = ℓXi = ℓ Xi = ℓNm where Nm = Xi
i =1 i =1 i =1

thus, Nm is the number of defaults in the portfolio up to time T

Since P [Lm = kℓ] = P [Nm = k], it is enough to study Nm .

Alexander Herbertsson (Univ. of Gothenburg) Financial Risk: Credit Risk, Lecture 2 April 27, 2017 4 / 29
 The mixed binomial model, cont.

Since X1 , X2 , . . . Xm are conditionally independent given Z , we have

 
m
P [ Nm = k | Z ] = p(Z )k (1 − p(Z ))m−k
k

Hence, we have
  
m
P [Nm = k] = E [P [ Nm = k | Z ]] = E p(Z )k (1 − p(Z ))k (5)
k
which holds regardless if Z is a discrete or continuous random variable.

If Z is a continuous random variable on R with density fZ (z) then

Z ∞  
m
P [Nm = k] = p(z)k (1 − p(z))m−k fZ (z)dz. (6)
−∞ k

We want to find the loss distribution FLm (x) = P [Lm ≤ x] for x ∈ [0, ∞), or
in fact for x ∈ [0, ℓ · m] (why ?)

Alexander Herbertsson (Univ. of Gothenburg) Financial Risk: Credit Risk, Lecture 2 April 27, 2017 5 / 29
 The loss distribution in a mixed binomial model
Note that for any positive x we have that
h xi h j x ki
FLm (x) = P [Lm ≤ x] = P [ℓNm ≤ x] = P Nm ≤ = P Nm ≤ (7)
ℓ ℓ
where ⌊y ⌋ is the integer part of y rounded downwards, e.g ⌊3.14⌋ = 3.
Pn
For n = 0, 1 . . . , m then P [Nm ≤ n] = k=0 P [Nm = k] which in (7) yields
⌊ ℓx ⌋
X
FLm (x) = P [Nm = k] (8)
k=0

where P [Nm = k] is computed by (5).

If Z is a continuous random variable on R with density fZ (z) then
P [Nm = k] is computed by (6) and this in (8) renders that
⌊ ℓx ⌋ Z ∞
 
X m
FLm (x) = p(z)k (1 − p(z))m−k fZ (z)dz. (9)
−∞ k
k=0

Note that FLm (x) in (8) or (9) will be piece-wise constant (i.e. flat) on each
interval [0, ℓ[, [ℓ, 2ℓ[. . . [(m − 1)ℓ, mℓ[, [mℓ, ∞[ (why ?)
Alexander Herbertsson (Univ. of Gothenburg) Financial Risk: Credit Risk, Lecture 2 April 27, 2017 6 / 29
 The loss distribution in a mixed binomial model, cont.
Note the formula for the loss distribution in (8) or (9) is rather tedious and
will fail for large values of m (why ?)

Fortunately, there is a very convenient approximation of the loss distribution

FLm (x) = P [Lm ≤ x] when m is ”large”

Recall that F (x) is the distrib. function of p(Z ), i.e F (x) = P [p(Z ) ≤ x]
and from last lecture we know that for any x ∈ [0, 1] it holds that
 
Nm
P ≤ x → F (x) = P [p(Z ) ≤ x] as m → ∞ (10)
m
We also have that
 
Nm x
FLm (x) = P [Lm ≤ x] = P [ℓNm ≤ x] = P ≤
m ℓm
and this in (10) then implies that
 x 
FLm (x) → F as m → ∞
ℓm
Alexander Herbertsson (Univ. of Gothenburg) Financial Risk: Credit Risk, Lecture 2 April 27, 2017 7 / 29
 The loss distribution in a mixed binomial model, cont.

Hence, if m is ”large” we have the following approximation for the loss

distribution FLm (x) = P [Lm ≤ x]
 x 
FLm (x) ≈ F if m is ”large”. (11)
ℓm
for any x ∈ [0, ℓm] and where F (x) = P [p(Z ) ≤ x].

x


So if m is ”large” we can approximate FLm (x) = P [Lm ≤ x] with F ℓm
instead of numerically compute the involved expression in the RHS of (9)

This will be very useful when computing different risk measures for credit
portfolios, such as Value-at-Risk and expected shortfall

Let us define/recap the concept of Value-at-Risk and expected shortfall

Alexander Herbertsson (Univ. of Gothenburg) Financial Risk: Credit Risk, Lecture 2 April 27, 2017 8 / 29
 Value-at-Risk
We now define/recap the risk measure Value-at-Risk, abbreviated VaR and
the below definition holds for any type of loss L (loss for equity risk, loss for
credit risk, loss operational risk etc etc)

Definition of Value-at-Risk
Given a loss L and a confidence level α ∈ (0, 1), then VaRα (L) is given by the
smallest number y such that the probability that the loss L exceeds y is no larger
than 1 − α, that is

VaRα (L) = inf {y ∈ R : P [L > y ] ≤ 1 − α}

= inf {y ∈ R : 1 − P [L ≤ y ] ≤ 1 − α}
= inf {y ∈ R : FL (y ) ≥ α}

where FL (x) is the distribution of L.

Linearity of Value-at-Risk (VaR): Let L be a loss and let a > 0 and b ∈ R be

constants. Then
VaRα (aL + b) = aVaRα (L) + b (12)
Alexander Herbertsson (Univ. of Gothenburg) Financial Risk: Credit Risk, Lecture 2 April 27, 2017 9 / 29
 Example of Value-at-Risk when L is continuous r.v.

Density of L

α = P [L ≤ VaRα (L)]

1 − α = P [L > VaRα (L)]

VaRα (L)

Figure: Visualization of definition of VaRα (L) when L is a continuous random

variable. The red region has the area 1 − α
Alexander Herbertsson (Univ. of Gothenburg) Financial Risk: Credit Risk, Lecture 2 April 27, 2017 10 / 29
 Value-at-Risk, cont.
Note that Value-at-Risk is defined for a fixed time horizon, so the above
definition should also come with a time period, e.g, if the loss L is over one
day, then we talk about a one-day VaRα (L).

In market risk, typically the underlying period studied for the loss is 1 day or
10 days.

In credit risk and in operational risk, one typically consider VaRα (L) for the
loss over one year.

Typical values for α are 95%, 99 or 99.9%, that is α = 0.95, α = 0.99 or

α = 0.999

Note that VaR, by definition, does not give any information about ”how bad
things can get”, i.e. the severity of the loss L which may occur with
probabilitiy 1 − α

We will later shortly discuss the expected shortfall which is a measure that
captures the severity of the loss L, given that L > VaRα (L).
Alexander Herbertsson (Univ. of Gothenburg) Financial Risk: Credit Risk, Lecture 2 April 27, 2017 11 / 29
 Value-at-Risk, cont.

Hence, by definition, VaRα (L) for a period T have the following

interpretation: ”We are α % certain that our loss L will not be bigger than
VaRα (L) dollars up to time T ”

However, we should keep in mind that this sentence can be very misleading
for several reasons.

One major reason is that VaRα (L) is computed under an assumption of how
the loss will be distributed, i.e. we use a specific model for L, and this
naturally leads to model risk

One typical example of model risk when computing VaRα (L) is that
FL (x) = P [L ≤ x] is assumed to have a distribution, which maybe (most
likely) not will match the ”true” distribution of L, which obviously is difficult
to know for sure.

Alexander Herbertsson (Univ. of Gothenburg) Financial Risk: Credit Risk, Lecture 2 April 27, 2017 12 / 29
 Inverse and generalized inverse functions
Recall that a function f (x) is strictly monotonic if it is strictly increasing or
strictly decreasing

Recall from your first year calculus course, that a strictly monotonic function
f (x) has a unique and well defined inverse f −1 (x) such that
1. f −1 (f (x)) = x , for all x in f-s domain
−1
1. f (f (y )) = y , for all y in f-s range
If the function f (x) is monotonic (i.e. not strictly monotonic) then the
concept of a inverse function has to be readjusted

Let us here focus on a nondecreasing function F (x).

Since F (x) is nondecreasing, it may be ”flat” for some regions in its domain
(see e.g. example on bottom on slide 6)

This means that in these ”flat” regions we can no longer find a unique
inverse function to F (x), so the concept of an inverse function must here be
redefined. Let us do this.
Alexander Herbertsson (Univ. of Gothenburg) Financial Risk: Credit Risk, Lecture 2 April 27, 2017 13 / 29
 Inverse and generalized inverse functions, cont

Definition of generalized inverse for a nondecreasing function

Let F (x) be a nondecreasing function on R, i.e. F (x) : R → R. The generalized
inverse F ← to F is then defined as

F ← (y ) = inf {x ∈ R : F (x) ≥ y } (13)

with the convention that inf of the empty set is ∞, i.e inf ∅ = ∞.

Note that if F (x) is a strictly increasing function then F ← = F −1 , that is

the generalized inverse F ← (y ) will simply be the ”usual” inverse F −1 (y )
defined as on the previous slides

By using the generalized inverse we can now define the α-quantile qα (F ) to a

distribution function F (x) as

qα (F ) = F ← (α) = inf {x ∈ R : F (x) ≥ α} , 0 < α < 1. (14)

Alexander Herbertsson (Univ. of Gothenburg) Financial Risk: Credit Risk, Lecture 2 April 27, 2017 14 / 29
 Generalized inverse, α-quantile and VaR

Hence, in view of the definition of a α-quantile (as a generalized inverse) qα (F )

and the definition of Value-at-Risk VaRα (L) we conclude that:

Value-at-Risk VaRα (L) is the α-quantile qα (FL ) of the loss distribution

FL (x) where FL (x) = P [L ≤ x], that is
VaRα (L) = FL← (α) = qα (FL ) (15)

In the case when FL (x) = P [L ≤ x] is continuous, and thus strictly increasing (i.e.
the loss L is a continuous random variable), FL (x) will not have any ”flat”
regions, so that FL← will be the usual inverse function FL−1 , and we then have that

VaRα (L) = FL−1 (α) = qα (FL ) (16)

Hence, if we can find an analytical expression for the inverse function FL−1 (y ), we
can then due to (16) also find an analytical expression for the risk-measure
Value-at-Risk VaRα (L)

Alexander Herbertsson (Univ. of Gothenburg) Financial Risk: Credit Risk, Lecture 2 April 27, 2017 15 / 29
 Value-at-Risk when L is a continuous random variable

If the loss L is a continuous random variable so that FL (x) is strictly

increasing and continuous, we have that FL−1 (y ) is also continuous, and thus
well defined and by definition

Furthermore, from the definition of an inverse function (see previous slides)

we have that FL (FL−1 (y )) = y for all y such that 0 < y < 1.

From (16) we have

VaRα (L) = FL−1 (α) (17)
so we then conclude that
FL (VaRα (L)) = FL (FL−1 (α)) = α (18)
that is,
FL (VaRα (L)) = α (19)
or alternatively,
P [L ≤ VaRα (L)] = α (20)

Alexander Herbertsson (Univ. of Gothenburg) Financial Risk: Credit Risk, Lecture 2 April 27, 2017 16 / 29
 Example of Value-at-Risk when L is continuous r.v.

1
α
Rx
FL (x) = f (y )dy
−∞ L

FL−1 (α) = VaRα (L)

fL (x)
α = P [L ≤ VaRα (L)]

1 − α = P [L > VaRα (L)]

VaRα (L)

Figure: Visualization of definition of VaRα (L) when L is a continuous random

variable. The red region has the area 1 − α
Alexander Herbertsson (Univ. of Gothenburg) Financial Risk: Credit Risk, Lecture 2 April 27, 2017 17 / 29
 Value-at-Risk for static credit portfolios

Consider mixed binomial model with m obligors and individual credit loss ℓ.
By linearity of VaR, see Equation (12), we can w.l.o.g assume that the size
of each loan is one monetary unit and that the loss ℓ is in %
Let F (x) = P [p(Z ) ≤ x] where p(Z ) is the mixing distribution where Z can
be a discrete or continuous random variable
If we use the exact loss distribution FLm (x) in (8) or (9) we compute VaR
via the generalized inverse of FLm (x)
However, if m is ”large” and Z is a continuous random variable so that F (x)
and F −1 (x) are continuous, we combine Equation (11) and (16) to get
VaRα (L) ≈ ℓ · m · F −1 (α) (21)

If m is ”large” and Z is a discrete random variable we combine Equation

(11) and (15) to get that
VaRα (L) ≈ ℓ · m · F ← (α) (22)
where F ← (x) is the generalized inverse of F (x) = P [p(Z ) ≤ x].
Alexander Herbertsson (Univ. of Gothenburg) Financial Risk: Credit Risk, Lecture 2 April 27, 2017 18 / 29
 Expected shortfall
The expected shortfall ESα (L) is defined as
1
1
Z
ESα (L) = VaRu (L)du.
1−α α

and if L is a continuous random variable one can show that

ESα (L) = E [ L | L ≥ VaRα (L)]

Let F (x) = P [p(Z ) ≤ x] where p(Z ) is the mixing distribution and Z is a

continuous random variable so that F (x) and F −1 (x) are continuous,

Hence, for the same static credit portfolio as on the two previous slides, when m
is large we have the following approximation formula for ESα (L)
1
ℓ·m
Z
ESα (L) ≈ F −1 (u)du
1−α α

Alexander Herbertsson (Univ. of Gothenburg) Financial Risk: Credit Risk, Lecture 2 April 27, 2017 19 / 29
 Mixed binomial models: the beta distribution
One example of a mixing binomial model is to let p(Z ) = Z where Z is a
beta distribution, Z ∼ Beta(a, b), which can generate heavy tails.
We say that a random variable Z has beta distribution, Z ∼ Beta(a, b), with
parameters a and b, if it’s density fZ (z) is given by
1
fZ (z) = z a−1 (1 − z)b−1 a, b > 0, 0<z <1 (23)
β(a, b)
where
1
Γ(a)Γ(b)
Z
β(a, b) = z a−1 (1 − z)b−1 dz = . (24)
0 Γ(a + b)
Here Γ(y ) is the Gamma function defined as
Z ∞
Γ(y ) = t y−1 e −t dt (25)
0
which satisfies the relation
Γ(y + 1) = y Γ(y ) (26)
for any y .
Alexander Herbertsson (Univ. of Gothenburg) Financial Risk: Credit Risk, Lecture 2 April 27, 2017 20 / 29
 Mixed binomial models: the beta distribution, cont.
By using Equation (24) and (26) one can show that β(a, b) satisfies the
recursive relation
a
β(a + 1, b) = β(a, b).
a+b
Also note that (23) implies that P [0 ≤ Z ≤ 1] = 1, that is Z ∈ [0, 1] with
probability one.
If Z has beta distribution with parameters a and b, then by using Equation
(24) and (26) one can show that
a a(a + 1)
and E Z 2 =
 
E [Z ] =
a+b (a + b)(a + b + 1)
so the above equations together with definition of Var(Z ) implies that
Var(Z ) = (a+b)2ab
(a+b+1) .

By varying the parameters a and b, the density fZ (z) can take on quite
different shapes (see next slide). Recall that fZ (z) is given by
1
fZ (z) = z a−1 (1 − z)b−1 a, b > 0, 0<z <1
β(a, b)
Alexander Herbertsson (Univ. of Gothenburg) Financial Risk: Credit Risk, Lecture 2 April 27, 2017 21 / 29
 Mixed binomial models: the beta distribution, cont.

Two different beta densities

14

a=1,b=9
12 a=10,b=90

10

8
density

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
x

Alexander Herbertsson (Univ. of Gothenburg) Financial Risk: Credit Risk, Lecture 2 April 27, 2017 22 / 29
 Mixed binomial models: the beta distribution, cont.

Consider a mixed binomial model where p(Z ) = Z has beta distribution with
parameters a and b. Then, by using (6) one can show that
 
m β(a + k, b + m − k)
P [Nm = k] = . (27)
k β(a, b)

It is possible to create heavy tails in the distribution P [Nm = k] by

choosing the parameters a and b properly in (27). This will then imply more
realistic probabilities for extreme loss scenarios, compared with the standard
binomial loss distribution (see figure on next page).
Furthermore, since p(Z ) = Z , the distribution of Nmm converges to the
distribution of the beta distribution, i.e
  Z x
Nm 1
P ≤x → z a−1 (1 − z)b−1 dz as m → ∞ (28)
m β(a, b) 0

and for large m we use (28) instead of the exact method via (27).

Alexander Herbertsson (Univ. of Gothenburg) Financial Risk: Credit Risk, Lecture 2 April 27, 2017 23 / 29
 Mixed binomial models: the beta distribution, cont.

The portfolio credit loss distribution in the standar and mixed binomial model
0.2

mixed binomial 50, beta(1,9)

0.18
binomial(50,0.1)

0.16

0.14

0.12
probability

0.1

0.08

0.06

0.04

0.02

0
0 5 10 15 20 25 30
number of defaults

Alexander Herbertsson (Univ. of Gothenburg) Financial Risk: Credit Risk, Lecture 2 April 27, 2017 24 / 29
 Mixed binomial models: logit-normal distribution
Another possibility for mixing distribution p(Z ) is to let p(Z ) be a
logit-normal distribution. This means that
1
p(Z ) =
1 + exp (−(µ + σZ ))
where σ > 0 and Z ∼ N(0, 1), that is Z is a standard normal random
variable. Note that p(Z ) ∈ [0, 1].
Furthermore, if x ∈ (0, 1) then p −1 (x) is well defined and given by
   
1 x
p −1 (x) = ln −µ . (29)
σ 1−x
 
The mixing distribution F (x) = P [p(Z ) ≤ x] = P Z ≤ p −1 (x) for a
logit-normal distribution is then given by
Z p−1 (x)
  1 z2
F (x) = P Z ≤ p −1 (x) = √ e − 2 dz = N(p −1 (x))
−∞ 2π
where p −1 (x) is given as in Equation (29) and N(x) is the distribution
function of a standard normal distribution.
Alexander Herbertsson (Univ. of Gothenburg) Financial Risk: Credit Risk, Lecture 2 April 27, 2017 25 / 29
 Mixed binomial models: logit-normal distribution, cont.

Furthermore, the distribution of Nmm converges to N(p −1 (x)), that is

 
Nm
P ≤ x → N(p −1 (x)) as m → ∞ (30)
m
where x ∈ (0, 1) and p −1 (x) is given as in Equation (29).
In a mixed binomial model with logit-normal distribution as above, it is
difficult to find closed formulas for quantities such as

P [Xi = 1] = E [p(Z )],

Var(Xi ) = E [p(Z )] (1 − E [p(Z )])
Cov(Xi , Xj ) = E p(Z )2 − E [p(Z )]2 = Var(p(Z )) for i 6= j
 

Hence, in the mixed binomial model with logit-normal distribution, the

above quantities have to be determined with a computer

Next lecture we will study a third mixed binomial model inspired by the
Merton model.
Alexander Herbertsson (Univ. of Gothenburg) Financial Risk: Credit Risk, Lecture 2 April 27, 2017 26 / 29
 Correlations in mixed binomial models

Recall the definition of the correlation Corr (X , Y ) between two random

variables X and Y , given by
Cov (X , Y )
Corr (X , Y ) = p p
Var (X ) Var (Y )

where Cov(X , Y ) = E [XY ] − E [X ] E [Y ] and Var (X )) = E X 2 − E [X ]2 .

 

Furthermore, also recall that Corr (X , Y ) may sometimes be seen as a

measure of the ”dependence” between the two random variables X and Y .

Now, let us consider a mixed binomial model as presented previously.

We are interested in finding Corr (Xi , Xj ) for two pairs i, j in the portfolio
(by the homogeneous-portfolio assumption this quantity is the same for any
pair i, j in the portfolio where i 6= j).

Below, we will therefore for notational convenience simply write ρX for the
correlation Corr (Xi , Xj ).
Alexander Herbertsson (Univ. of Gothenburg) Financial Risk: Credit Risk, Lecture 2 April 27, 2017 27 / 29
 Correlations in mixed binomial models, cont.

Recall from previous slides that P [ Xi = 1 | Z ] = p(Z ) where p(Z ) is the

mixing variable.
Furthermore, we also now that

Cov(Xi , Xj ) = E p(Z )2 − p̄ 2
 
and Var(Xi ) = p̄(1 − p̄) (31)

where p̄ = E [p(Z )].

Thus, the correlation ρX in a mixed binomial models is then given by

E p(Z )2 − p̄ 2
 
ρX = (32)
p̄(1 − p̄)

where p̄ = E [p(Z )] = P [Xi = 1] is the default probability for each obligor.

Hence, the correlation ρX in a mixed binomial is completely determined by
the fist two
 moments of the mixing variable p(Z ), that is E [p(Z )] and
E p(Z )2 .

Alexander Herbertsson (Univ. of Gothenburg) Financial Risk: Credit Risk, Lecture 2 April 27, 2017 28 / 29
 Thank you for your attention!

Alexander Herbertsson (Univ. of Gothenburg) Financial Risk: Credit Risk, Lecture 2 April 27, 2017 29 / 29