ACTSC 445/845: Quantitative Risk Management

Fall 2022

Erik Hintz
Department of Statistics and Actuarial Science
[email protected]

Lecture 02

1 / 22
 Today’s Agenda

Last time:

Who’s this guy and syllabus

What is Risk Management?
I Risk, Types of Risk
I History of RM
I Path to regulation and Basel’s 3 Pillar Concept
I The nature of the challenge

Today (Lec 2, 09/13):

Risk Management for a financial firm

Risk factors and loss distributions

2 / 22
Chapter 2: Basic Concepts in Risk Management

2.1. Risk Management for a Financial Firm

3 / 22
 Balance Sheet of a Bank

Simplified balance sheet of a bank:

Assets Liabilities
Cash 10M Customer deposits 80M
Securities 50M Bonds issued 40M
(bonds, stocks, derivatives)
Loans and Mortgages 100M Short Term Borrowing 30M
- corporates
- retail and smaller clients Reserves (for losses on loans) 20M
- government
Other Assets 20M Debt (sum of the above) 170M
- property
- investment in companies
Short-term lending 20M Equity 30M
Total 200M Total 200M

4 / 22
 Risks faced by a financial firm

Decrease in the value of the investments on the asset side

(market risk, credit risk)
Maturity Mismatch: Assets often illiquid (long-term) but liabilities often
short-term obligations
⇒ Solvent bank can default!
Risk for insurer: insolvency (claims of policyholders cannot be met)
Asset side: Similar to bank (same risks).
Liability side: Risk that reserves are insufficient.
For a life insurer, liabilities are of long-term nature. More types of risk
relevant here (interest rate risk, inflation risk, longevity risk)
Valuation of the items on the balance sheet non-trivial!

5 / 22
 Different Notions of Capital

The following are some notions of capital:

Equity Capital:
Value of Assets - Debts.
Measures the firm’s value to its shareholders.
Regulatory capital:
Capital required according to regulatory rules.
For European insurance firms: Minimum (MCR) and Solvency Capital
Requirements (SCR).
Economic capital:
Capital required to control the probability of becoming insolvent (typically over
one year).
Internal assessment of (total/overall) risk capital.
Aims at a holistic view (assets and liabilities) and works with fair values of
balance sheet items
In what follows, we assume the correct notion of capital (and other quantities)
and treat these numbers as given.

6 / 22
Chapter 2: Basic Concepts in Risk Management

2.2. Risk Factors and Loss Distributions

7 / 22
 Mapping of Risks: General definitions

Goal: General probabilistic framework for modelling financial risk.

We work on a probability space denoted by (Ω, F , P).
A risk or loss is then a random variable (rv) L : Ω → R and we denote by
FL (x ) = P(L ≤ x ) its (cumulative) distribution function ((c)df).
Consider a portfolio of assets/liabilities. Then

Vt = value of portfolio at time t

Vt is a random variable and assumed to be known at time t (today).

Interest lies in the loss of the portfolio over a time horizon ∆t (eg 1/250).
Assume
I composition of portfolio unchanged over ∆t.
I no payments during ∆t.
Fair for small ∆t, debatable for large ∆t.
For simplicity, we often write t + 1 instead of t + ∆t (think of everything
being measured as multiples of ∆t).

8 / 22
 General definitions

The change in value of the portfolio is given by

∆Vt +1 = Vt +1 − Vt

The (random) loss is the sign adjusted change in value:

Lt +1 = −∆Vt +1 = −(Vt +1 − Vt )

The distribution of Lt +1 is called loss distribution.

If ∆t is large, should consider discounted values, i.e.
Vt +1
∆Vt +1 = − Vt
1+r
where r is the risk free interest rate over the time period ∆t, possibly
random. We mostly ignore this issue.

9 / 22
 Modelling Vt
We model the rv Vt as a function of time (t) and a d −dimensional
random vector Z t = (Zt,1 , . . . , Zt,d ) (known at time t) of risk factors:

Vt = f (t, Z t )

for a measurable function f : R+ × Rd → R.

Choice of f and Z t is a modelling issue and problem-specific.
Define risk factor changes as

X t +1 = Z t +1 − Z t

Then we can write

Lt +1 = −(f (t + 1, Z t +1 ) − f (t, Z t ))
= − (f (t + 1, Z t + X t +1 ) − f (t, Z t ))
⇒ At time t, distribution of Lt +1 is determined by the distribution of
X t +1 (typically not linear in X t +1 ⇒ cdf may be hard to determine)
With L(x ) = − (f (t + 1, Z t + x ) − f (t, Z t )) (loss operator) we can write

Lt + 1 = L ( X t + 1 )
10 / 22
 Linearized Loss

Goal: Find linear approximation to Lt +1 (as function of X t +1 ) that is

easier to work with.
Assume f is differentiable. A first order Taylor approximation yields
d
f (t + 1, Z t + X t +1 ) ≈ f (t, Z t ) + ft (t, Z t ) · 1
|{z} + ∑ fz (t, Z t ) · Xt +1,j
j

=(t +1)−t j =1

∂f ∂f
where ft = ∂t and fzj = ∂zj .
Can then approximate Lt +1 by Lt∆+1 , the linearized loss:
 d 
Lt∆+1 = − ft (t, Z t ) + ∑ fzj (t, Z t ) ·Xt +1,j = −(ct + b t> X t +1 ) (1)
| {z } j =1 | {z }
=:ct =:bt,j

where b t = (bt,1 , . . . , bt,d ).

Lt∆+1 is a linear function of X t +1 = (Xt +1,1 , . . . , Xt +1,d ).
Approximation best if ∆t small, if risk factor changes X small in absolute
value and if Vt depends almost linearly on the risk factors (ie if f has
small second derivative)

11 / 22
Summary: Concept Map

12 / 22
Example (Stock portfolio)
Given: Portfolio of d stocks St,1 , . . . , St,d where

St,j = value of jth stock at time t

and λj is the number of shares of stock j.

Consider logarithmic prices as risk factors, that is

Zt,j = log St,j , j = 1, . . . , d

The value of the portfolio at time t is then

d d
Vt = f (t, Z t ) = ∑ λj St,j = ∑ λj e Z t,j

j =1 j =1

so that our mapping f : R+ × Rd → R is given by

d
f (t, z ) = ∑ λj e z
j
where z = (z1 , . . . , zd )>
j =1

13 / 22
One period ahead loss (assuming ∆t = 1) is then

Lt +1 = −(Vt +1 − Vt ) = − (f (t + 1, Z t + X t +1 ) − f (t, Z t ))
d   d  
= − ∑ λj e Zt,j +Xt +1,j − e Zt,j = − ∑ λj e Zt,j e Xt +1,j − 1
j =1 j =1
d  
=− ∑ |λj{z
St,j e Xt +1,j − 1
j =1 }
=:wt,j

which is non-linear in X t +1 .
The loss operator is given by
d
L(x ) = − ∑ wt,j (e x j
− 1) , x = (x1 , . . . , xd )
j =1

14 / 22
The linearized loss is given by
 d   d 
Lt∆+1 = − ft (t, Z t ) + ∑ fz (t, Z t ) · Xt +1,j
j = − 0+ ∑ wt,j Xt +1,j
j =1 j =1

= −w t> X t +1

So
Lt∆+1 = −w t> X t +1
which has the form Lt∆+1 = −(ct + b t> X t +1 ) (as in (1)) for ct = 0 and
bt = w t

15 / 22
 Why is the linear approximation Lt∆+1 useful?

The linear approximation Lt∆+1 can be used to find an approximation for

the expectation and variance of Lt +1 :
Assume µ = E (X t +1 ) and Σ = Cov(X t +1 ) both exist and are known.
Then, expectation and variance of the (linearized) one-period ahead loss
are

E Lt∆+1 = E −w t> X t +1 = −w t> µ

   

Var Lt∆+1 = Var −w t> X t +1 = w t> Cov (X t +1 ) w t = w t> Σw t

   

If we further assume that X t +1 is multivariate normal (i.e.

X t +1 ∼ Nd (µ, Σ)), then Lt∆+1 is also univariate normal:

Lt∆+1 ∼ N −w t> µ, w t> Σw t

 

The (actual) loss Lt +1 will not be normal, even if X t +1 is assumed to be

normal.

16 / 22
 Loss Distributions

So far: Determined a mapping f that maps risk factors to (changes in) values
or losses.
The following key statistical tasks of QRM now need to be tackled:
1) Find a statistical model for X t +1 .
Typically model for forecasting, estimated using historical data.
2) Compute/derive/approximate the cdf FLt +1 of Lt +1 .
This requires the cdf of f (t + 1, Z t + X t +1 ).
3) Compute a risk measure (later) from FLt +1 .

In 2) and 3) will sometimes need to use linearized loss Lt∆+1 with cdf FL∆ as
t +1
approximation to Lt +1 with cdf FLt +1 .
We discuss three general methods to approach these challenges: Analytical
method, Historical simulation and Monte Carlo.

17 / 22
 Method 1: Analytical Method

Main idea: Choose FX t +1 and f such that FLt +1 can be determined explicitly.

We illustrate the variance-covariance method. We assume

1) X t +1 ∼ Nd (µ, Σ) (eg if (Z t ) Brownian Motion (BM), (S t ) geometric
BM)
2) FL∆ is a good approximation to FLt +1
t +1

Then by Eq. (1)

Lt∆+1 = −(ct + b t> X t +1 ) =⇒ Lt∆+1 ∼ N(−ct − b t> µ, b t> Σb t )

Ass. 1

where we recall that ct and b t are known at time t. The vector µ and the
matrix Σ need to be estimated (e.g. via time-series), if not available.

18 / 22
 Method 1: Analytical Method

Advantages Drawbacks
Lt∆+1 may be a poor
FL∆ explicit and thus (usually) approximation to Lt +1
t +1
risk measures explicit as well Ass. 1 unrealistic: Stylized facts
about X t +1 suggest FX t +1 has a
thinner body/heavier tail than
Easy to implement (if d is not Nd (µ, Σ) ⇒ Ass. 1 yields to
too large) underestimation of the tail of
FLt +1 and thus risk measures.

We remark that Ass. 1 (and thus the second drawback) can be relaxed by
considering other multivariate distribution families that are also closed under
linear operations, e.g. the multivariate t distribution. See later.

19 / 22
 Method 2: Historical simulation

Main idea: Estimate FLt +1 by its empirical distribution function (edf).

Given n past risk factor changes X t −n+1 , . . . , X t , set

Lk = L(X k ) = − (f (t + 1, Z t + X k ) − f (t, Z t )) , k ∈ {t − n + 1, . . . , t } (2)

Interpretation: Lk shows what would happen to our current portfolio if the

past risk factor change X k were to recur.
FLt +1 can then be estimated via

1 n
FbLt +1 ,n (x ) = ∑ 1{Lt −i +1 ≤x } , x ∈ R (3)
n i =1

and all further calculations (eg risk measures) can be based on FbLt +1 ,n (x ) as an
approximation to FLt +1 .
The same logic holds if one wishes to use Lt∆+1 as an approximation to Lt +1 , in
which case one uses FbL∆ as an approximation to FLt +1 .
t +1

20 / 22
 Method 2: Historical simulation

Advantages Drawbacks
Easy to implement
(one-dimensional problem) Sufficient data for all risk-factor
No estimation of the distribution changes required
of X t +1 needed
No assumptions on the Only past losses considered
dependence structure of risk (”driving a car by looking in the
factor changes necessary back mirror”)

21 / 22
 Method 3: Monte Carlo

Main idea: Take any model for X t +1 , simulate X t +1 , compute corresponding

losses Lk as in (2) and estimate FLt +1 (typically via FbLt +1 ,n (x ) as in (3))

Advantages Drawbacks
Unclear how to find an
Quite general (applicable to any
appropriate model for X t +1
model which we can sample
from) Computational costs can be high
(every simulation requires
Quite flexible (can easily change evaluation of f , possibly
parameters and/or distributions) expensive, as in nested Monte
Carlo simulations)

22 / 22