Lecture Note of Bus 41202, Spring 2009:

Value at Risk & Risk Management

Classification of Financial Risk

1. Credit risk

2. Market risk

3. Operational risk

We start with the market risk, because

• more data are available

• easier to understand

• the idea applicable to other types of risk.

What is Value at Risk (VaR)?

• a measure of minimum loss of a given asset

• amount a position could decline in a given period, associated

with a given probability (or confidence level)

A formal definition:

• time period given: ∆t = `

• change in value: ∆V (`)

• CDF of the change F`(x)

1
 • given (tail) probability: p

• a long position:

p = P r[∆V (`) ≤ VaR] = F`(VaR).

• a short position:

p = P r[∆V (`) ≥ VaR]

= 1 − P r[∆V (`) ≤ VaR]
= 1 − F`(VaR).

We can study VaR by simply focusing on the loss function.

Quantile: xp is the pth quantile of F`(x) if

p = F`(xp)

and F`(.) is continuous.

Factors affect VaR:

1. the probability p.

2. the time horizon `.

3. data frequency.

4. the CDF F`(x). (or CDF of loss)

5. the mark-to-market value of the position.

2
Why use log returns?
log returns ≈ percentage changes.
VaR = Value × (VaR of log return).

Methods available for market risk

1. RiskMetrics

2. Econometric modeling

3. Empirical quantile

4. Traditional extreme value theory (EVT)

5. EVT based on exceedance over a high threshold

Data used in illustrations:

Daily log returns of IBM stock

• span: July 3, 62 to Dec. 31, 98.

• size: 9190 points

Position: long on $10 millions.

RiskMetrics

• Developed by J.P. Morgan

• rt given Ft−1: N (0, σt2)

• σt2 follows the special IGARCH(1,1) model

σt2 = ασt−1
2 2
+ (1 − α)rt−1 , 1 > α > 0.

3
 • VaR = 1.65σt if p = 0.05.
√
• k-horizon: VaR[k] = kVaR
The square root of time rule

• Pros: simplicity and transparence

• Cons: model is not adequate

Example: IBM data

Model:

rt = at, at = σtt,
2
σt2 = 0.9396σt−1 + (1 − 0.9396)a2t−1
2
Because r9190 = −0.0128 and σ̂9190 = 0.0003472,
2
σ̂9190 (1) = 0.000336.
√
For p = 0.05, VaR of rt = −1.65 × 0.000336 = −0.03025

VaR = $10, 000, 000 × 0.03025 = $302, 500.

√
For p = 0.01, VaR of rt = −2.3262 × 0.000336 = −0.04265, and
VaR = $426,500.

Econometric models

• rt = µt + at given Ft−1

• µt: a mean equation (Ch.2)

• σt2: a volatility model (Ch. 3 or 4)

4
 • Pros: sound theory

• Cons: a bit complicated.

IBM data:
Case 1: Gaussian

rt = 0.00066 − 0.0247rt−2 + at, at = σtt

σt2 = 0.00000389 + 0.0799a2t−1 + 0.9073σt2.

2
From r9189 = −0.00201, r9190 = −0.0128 and σ9190 = 0.00033455,
we have
2
r̂9190(1) = 0.00071 and σ̂9190 (1) = 0.0003211.

If p = 0.05, then
√
0.00071 − 1.6449 × 0.0003211 = −0.02877.

VaR = $10,000,000×0.02877 = $287,700.

If p = 0.01, then the quantile is
√
0.00071 − 2.3262 × 0.0003211 = −0.0409738.

VaR = $409,738.

Case 2: Student-t5

rt = 0.0003 − 0.0335rt−2 + at, at = σtt

σt2 = 0.000003 + 0.0559a2t−1 + 0.9350σt−1

2
.
5
 2
From the data, r9189 = −0.00201, r9190 = −0.0128 and σ9190 =
0.000349, we have
2
r̂9190(1) = 0.000367 and σ̂9190 (1) = 0.0003386.

If p = 0.05, then the quantile is

r √ √
.000367−(2.015/ 5/3) .0003386 = .000367−1.5608 .0003386 = −.028354.

VaR = $10, 000, 000 × 0.028352 = $283,520.

If p = 0.01, the quantile is
r √
0.000367 − (3.3649/ 5/3) 0.0003386 = −0.0475943.

VaR = $475,943.
Discussion:

• Effects of heavy-tails seen with p = 0.01.

• Multiple step-ahead forecasts are needed.

Example 7.3 (continued). 15-day horizon.

r̂9190[15] = 0.00998 and σt[15] = 0.0047948.
√
If p = 0.05, the quantile is 0.00998−1.6449 0.0047948 = −0.1039191.
15-day VaR = $10,000,000×0.1039191 = $1,039,191.
√
RiskMetrics: VaR = $287, 700 × 15 = $1,114,257.

Empirical quantile
Sample of log returns: {rt|t = 1, · · · , n}.
Order statistics:
6
 r(1) ≤ r(2) ≤ · · · ≤ r(n)

r(i) as the ith order statistic of the sample.

r(1) is the sample minimum
r(n) the sample maximum.
Idea: Use the empirical quantile to estimate the theoretical quantile
of rt.
For a given probability p, what is the empirical quantile?
If np = ` is an integer, then it is r(`).
If np is not an integer, find the two neighboring integers `1 < np < `2
and use interpolation.
The quantile is
p2 − p p − p1
x̂p = r(`1) + r(` ).
p 2 − p1 p 2 − p1 2
IBM data:
n = 9190. If p = 0.05, then np = 459.5.
5% quantile is (r(459) + r(460))/2 = −0.021603.
VaR = $216,030.
If p = 0.01, then np = 91.9 and the 1% quantile is
p2 − 0.01 0.01 − p1
x̂0.01 = r(91) + r(92)
p2 − p1 p2 − p 1
.00001 0.0001
= (−3.658) + (−3.657)
.00011 0.00011
≈ −3.65709.

VaR is $365,709.
7
Extreme value theory: Focus on the tail behavior of rt.
Review of extreme value theory
A properly normalized r(1) assumes a special distribution:

1 − exp[−(1 + kx)1/k ] if k 6= 0








F∗(x) = 
 

1 − exp[− exp(x)] if k = 0





for x < −1/k if k < 0 and for x > −1/k if k > 0.

k: the shape parameter
α = −1/k: tail index of the distribution.
Classification of distributions:
• Type I: k = 0, the Gumbel family. The CDF is
F∗(x) = 1 − exp[− exp(x)], −∞ < x < ∞. (1)

• Type II: k < 0, the Fréchet family. The CDF is




 1 − exp[−(1 + kx)1/k ] if x < −1/k
F∗(x) =  (2)
  1 otherwise.
• Type III: k > 0, the Weibull family. The CDF here is



 1 − exp[−(1 + kx)1/k ] if x > −1/k
F∗(x) = 
  0 otherwise.
The probability density function (pdf) of the normalized minimum
is 
(1 + kx)1/k−1 exp[−(1 + kx)1/k ] if k 6= 0








f∗(x) = 



exp[x − exp(x)] if k = 0





8
where −∞ < x < ∞ for k = 0, x < −1/k for k < 0 and x > −1/k
for k > 0.

How to use the EVT distribution?

If we know the three parameters, we can compute the quantile!
Empirical estimation
Divide the sample into non-overlapping subsamples.
Suppose there are T data points, we devide the data as

{r1, · · · , rn|rn+1, · · · , r2n|r2n+1, · · · , r3n| · · · |r(g−1)n+1, · · · , rng },

n: size of subgroup
Idea: find the minimum of each subgroup. These minima are the
data used to estimate the three parameters.
Several estimation methods available. We use maximum likelihood
estimates.
IBM data:

9
 n g Scale αn Location βn Shape Par. kn
(a) Minimal returns
21 437 0.823(0.035) −1.902(0.044) −0.197(0.036)
63 145 0.945(0.077) −2.583(0.090) −0.335(0.076)
126 72 1.147(0.131) −3.141(0.153) −0.330(0.101)
252 36 1.542(0.242) −3.761(0.285) −0.322(0.127)
(b) Maximal returns
21 437 0.931(0.039) 2.184(0.050) −0.168(0.036)
63 145 1.157(0.087) 3.012(0.108) −0.217(0.066)
126 72 1.292(0.158) 3.471(0.181) −0.349(0.130)
252 36 1.624(0.271) 4.475(0.325) −0.264(0.186)

EVT to VaR: Use a two-step procedure, because of the division into

subgroup.
VaR for rt:

kn
 
αn
βn − 1 − [−n ln(1 − p)] if kn 6= 0






 kn

VaR = 
 

βn + αn ln[−n ln(1 − p)] if kn = 0.





For IBM data, if n = 63 (quarterly minima), then α̂n = 0.945,

β̂n = −2.583, and k̂n = −0.335. If p = 0.01, the VaR is
0.945  −0.335

VaR = −2.583 − 1 − [−63 ln(1 − 0.01)]
−0.335
= −3.04969

VaR is $304,969.
10
If p = 0.05, then VaR is $166,641.
For n = 21, the results are:
VaR = $340,013 for p = 0.01;
VaR = $184,127 for p = 0.05.
Discussion:

• Results depend on the choice of n

• VaR seems low, but it might be due to the choice of p.

If p = 0.001, then
VaR = $546,641 for the Gaussian AR(2)-GARCH(1,1) model
VaR = $666,590 for the extreme value theory with n = 21.

Summary of IBM data:

Position = $10 millions.
If p = 0.05, then

1. $302,500 for the RiskMetrics,

2. $287,200 for an AR(2)-GARCH(1,1) model,

3. $283,520 for an AR(2)-GARCH(1,1) with t5

4. $216,030 using the empirical quantile, and

5. $184,127 for EVT with n = 21.

p = 0.01, then

1. $426,500 for the RiskMetrics,

11
 2. $409,738 for an AR(2)-GARCH(1,1) model,

3. $475,943 for an AR(2)-GARCH(1,1) model with t5

4. $365,800 for empirical quantile, and

5. $340,013 for EVT with n = 21.

If p = 0.001, then

1. $566,443 for the RiskMetrics,

2. $546,641 for an AR(2)-GARCH(1,1) model,

3. $836,341 for an AR(2)-GARCH(1,1) model with t5

4. $780,712 for quantile, and

5. $666,590 for EVT with n = 21.

Multi-period VaR with EVT

VaR(`) = `1/α VaR = `−k VaR

where α is the tail index and k is the shape parameter.

For IBM data with p = 0.05 and n = 21,

VaR(30) = (30)0.197VaR = 1.954 × $184, 127 = $359, 841.

12
New approach to VaR
Based on Exceedances over a high threshold
Idea: frequency of big returns and their magnitudes are important.
Statistical theory:
Two-dimensional Poisson process
Two possible cases:
Homogeneous: parameters are fixed over time
Non-homogeneous case: parameters are time-varying, according to
some explanatory variables.
IBM data: homogeneous model

Thr. Exc. Shape Par. k Log(Scale) ln(α) Location β

(a) Original log returns
3.0% 175 −0.30697(0.09015) 0.30699(0.12380) 4.69204(0.19058)
2.5% 310 −0.26418(0.06501) 0.31529(0.11277) 4.74062(0.18041)
2.0% 554 −0.18751(0.04394) 0.27655(0.09867) 4.81003(0.17209)
(b) Removing the sample mean
3.0% 184 −0.30516(0.08824) 0.30807(0.12395) 4.73804(0.19151)
2.5% 334 −0.28179(0.06737) 0.31968(0.12065) 4.76808(0.18533)
2.0% 590 −0.19260(0.04357) 0.27917(0.09913) 4.84859(0.17255)
VaR calculation:

β + αk 1 − [−T ln(1 − p)]k
 
if k 6= 0








VaR = 
 

β + α ln[−T ln(1 − p)] if k = 0





13
where T = 252, the number trading days in a year.
IBM data: VaR of 5% & 1%
• Case I: original returns

1. η = 3.0%: $228,239 & $359.303.

2. η = 2.5%: $219,106 & $361,119.
3. η = 2.0%: $212,981 & $368.552.

• Case II: remove sample mean

1. η = 3.0%: $232,094 & $363,697.

2. η = 2.5%: $225,782 & $364,254.
3. η = 2.0%: $217,740 & $372,372.

Non-homogeneous case:

kt = γ0 + γ1x1t + · · · + γv xvt ≡ γ0 + γ 0xt

ln(αt) = δ0 + δ1x1t + · · · + δv xvt ≡ δ0 + δ 0xt
βt = θ0 + θ1x1t + · · · + θv xvt ≡ θ0 + θ 0xt.

For IBM data, explanatory variables include past volatilities, etc.

See Chapter 7 for more details and estimation results.
Illustration:
For December 31, 1998, we have x3,9190 = 0, x4,9190 = 0.9737 and
x5,9190 = 1.9766. The parameters become

k9190 = −0.01195, ln(α9190) = 0.19331, β9190 = 6.105.

14
If p = 0.05, then quantile = 3.03756% and

VaR = $10, 000, 000 × 0.0303756 = $303, 756.

If p = 0.01, then Var is $497,425.

For December 30, 1998, we have x3,9189 = 1, x4,9189 = 0.9737 and

x5,9189 = 1.8757 and

k9189 = −0.2500, ln(α9189) = 0.52385, β9189 = 5.8834.

The 5% VaR becomes

VaR = $10, 000, 000 × 0.0269139 = $269, 139.

If p 0.01, then VaR becomes $448,323.

R and S-Plus Demonstration:

Both packages use the library: evir. Download the library before
using the commands.

Please note that the shape parameter "k" of chapter 7

is denoted by "minus xi" in evir.
Also, the program ‘‘evir’’ uses maxima (right tail) so that
one should use ‘‘minus returns’’ for the left tail.
(Command line starts with ‘‘> *’’)
(* Output is edited *)

(* pgev, dgev, qgev and rgev are commands for CDF, pdf, quantile and rando
generalized extreme value distribution *)
(* For example, to obtain the 95th quantile, use below *)

15
> qgev(0.95,xi=0.5,mu=0,sigma=1) % obtain quantile
[1] 6.830793

Example: daily log returns, in percentages, of IBM stock: 1962 to 1998.

> library(evir)
> da=read.table("d-ibmln98.dat")
> ibm=da[,1]
> plot(ibm,type=’l’)
> qqnorm(ibm) % normal probability plot

> nibm=-ibm % Focus on the left tail.

> m1=gev(nibm,block=21) % fit gen. extreme value dist.

> m1
$n.all
[1] 9190
$n
[1] 438
$data
[1]3.288 3.619 3.994 3.864 1.824 2.161 1.578 1.535 0.911 1.019
.....
[431] 3.502 2.501 4.576 8.456 3.723 1.082 2.910 3.099
$block
[1] 21

$par.ests
xi sigma mu
0.1956199 0.8239793 1.9031998
$par.ses
xi sigma mu

16
0.03554473 0.03476737 0.04413629

$varcov
[,1] [,2] [,3]
[1,] 1.263428e-03 -2.782725e-05 -0.0004338483
[2,] -2.782725e-05 1.208770e-03 0.0008475859
[3,] -4.338483e-04 8.475859e-04 0.0019480124

$converged
[1] 0
$nllh.final
[1] 654.3337
attr(,"class")
[1] "gev"

> names(m1)
[1] "n.all" "n" "data" "block" "par.ests"
[6] "par.ses" "varcov" "converged" "nllh.final"

> m1$n % numbers of monthly maximum

[1] 438
> ymax=m1$data
> hist(ymax)
> ysort=sort(ymax)
> plot(ysort,-log(-log(ppoints(ysort))),xlab=’Monthly maximum’)
% Gumbel qq-plot

> plot(m1) % Model chekcing plots

Make a plot selection (or 0 to exit):

17
1: plot: Scatterplot of Residuals
2: plot: QQplot of Residuals
Selection: 1

Make a plot selection (or 0 to exit):

1: plot: Scatterplot of Residuals
2: plot: QQplot of Residuals
Selection: 2

Make a plot selection (or 0 to exit):

1: plot: Scatterplot of Residuals
2: plot: QQplot of Residuals
Selection: 0

> 1-pgev(max(ymax),xi=.196,mu=1.90,sigma=.824)
[1] 5.857486e-05 % Prob. that the drop will exceed the maximum.

> rlevel.gev(m1,k.blocks=36) %return level & its 95% conf. interval.

[1] 5.568969 6.158516 6.980167

** Peak Over the Threshold approach (homogeneous case).

> meplot(nibm) % mean excess plot

> m2=gpd(nibm,threshold=2.5)
> names(m2)
[1] "n" "data" "threshold" "p.less.thresh"
[5] "n.exceed" "method" "par.ests" "par.ses"
[9] "varcov" "information" "converged" "nllh.final"

18
> m2$threshold
[1] 2.5
> m2$n.exceed
[1] 310

> m2 % Obtain all output

$n
[1] 9190
$data
[1] 3.288 2.649 2.817 3.619 3.994 3.795 3.784 2.623 3.864 3.149
.....
[301]2.501 2.835 4.576 4.393 8.456 2.916 3.723 2.910 2.654 3.099

$threshold
[1] 2.5
$p.less.thresh
[1] 0.9662677
$n.exceed
[1] 310
$method
[1] "ml"

$par.ests
xi beta
0.2641593 0.7786761

$par.ses
xi beta
0.06659234 0.06714131

19
$varcov
[,1] [,2]
[1,] 0.004434540 -0.002614442
[2,] -0.002614442 0.004507955
$information
[1] "observed"
$converged
[1] 0
$nllh.final
[1] 314.375
attr(,"class")
[1] "gpd"

> plot(m2) % Model checking. Should see all plots.

Make a plot selection (or 0 to exit):

1: plot: Excess Distribution
2: plot: Tail of Underlying Distribution
3: plot: Scatterplot of Residuals
4: plot: QQplot of Residuals
Selection: 1
[1] "threshold = 2.5 xi = 0.264 scale = 0.779 location= 2.5"

Make a plot selection (or 0 to exit):

1: plot: Excess Distribution
2: plot: Tail of Underlying Distribution
3: plot: Scatterplot of Residuals
4: plot: QQplot of Residuals

20
Selection: 0

> shape(nibm) % A plot showing the stability of the estimates.

> riskmeasures(m2,c(0.95,0.99)) % Compute VaR and expected shortfall.

p quantile sfall
[1,] 0.95 2.208932 3.162654
[2,] 0.99 3.616487 5.075507

Additional information on applying extreme value theory to value at

risk calculation.
To traditional approach of EVT
Return Level: It is a risk measure based on the idea of subperiods.
The g n-subperiod return level, Ln,g , is the level that is exceeded in
one out of every g subperiods of length n.
1
P (rn,i < Ln,g ) = ,
g
where n is the length of subperiod used in estimating the GEV model
and rn,i denotes subperiod minimum. For sufficiently large n,
αn
Ln,g = βn + {[− ln(1 − 1/g)]kn − 1},
kn
where the shape parameter kn 6= 0.
For a short position, the return level is
αn
Ln,g = βn + {1 − [− ln(1 − 1/g)]1/kn }.
kn

Peaks over the Threshold (POT)

21
Generalized Pareto Distribution: For simplicity, assume that
the shape parameter k 6= 0. Consider the extreme value distribution
of maximum (Eq. (7.29) of the textbook)
  1/k 
k(r − β)
F∗(r) = exp − 
1 − .
  
 
α


The distribution of r ≤ x + η given η, where x ≥ 0, is

 1/k
kx 
1 −
Pr(r ≤ x + η|r > η) ≈ 1 −   ,
ψ(η)
where ψ(η) = α − k(η − β), which depends on η.
The distribution with cumulative distribution function
 1/k
kx 
G(x) = 1 − 1 −  ,
ψ(η)
is called a generalized Pareto distribution (GPD).

Selection of the high threshold

Mean Excess: Given a high threshold ηo, suppose the excess r −ηo
follows a GPD with parameter k and ψ(ηo), where 0 > k > −1.
Then the mean excess over the threshold is
ψ(ηo)
E(r − ηo|r > ηo) = .
1+k
For any η > ηo, the mean excess function is defined as
ψ(ηo) − k(η − ηo)
e(η) = E(r − η|r > η) = .
1+k
The fact that, for a given k, e(η) is a linear function of η, where
η > ηo, provides a simple method to infer the threshold ηo for GPD.
22
Define the empirical mean excess as
1 N η
eT (η) = (rt − η),
X

Nη i=1 i
where Nη is the number of returns that exceed η and rti are the
values of the corresponding returns.
The scatterplot eT (η) versus η is called the mean excess plot, which
should be linear for η > ηo.
In R or S-Plus, the command is meplot.

Use of GPD in VaR

For a given threshold, estimate GPD to obtain parameters k and
ψ(η). Check the adequacy of the fit; see demonstration. Provided
that the model is adequate, the VaR can be computed by
  k 
ψ(η) 


 T



VaRq = η + 1 −  (1 − q) 
 ,
k  
 Nη 



where q = 1 − p with 0 < p < 0.05, T is the sample size and Nη is

the number of exceedances.
Alternatively, one can use the formula in Eq. (7.38) of the text-
book when one treats the exceedances and exceeding times as a two-
dimensional Poisson process. The VaR results obtained are close.

Expected Shortfall (ES): the expected loss given that the VaR
is exceeded. Specifically,

EESq = E(r|r > VaRq ) = VaRq + E(r − VaRq |r > VaRq ).

23
For GPD, it turns out that
VaRq ψ(η) + kη
ESq = + .
1+k 1+k
In evir, the command is riskmeasures.

Let rn,i be the maximum of a subperiod of length n. Under the

traditional EVT, rn,i follows a generalized extreme value distribution
with parameter (xi,sigma,mu).
What is the relationship between quantile of rn,i and the return rt?
Let Q be a real number.

P (rn,i > Q) = 1 − P (rn,i ≤ Q)

= 1 − P (all rt in the subperiod ≤ Q)
n
= 1− P (rt ≤ Q) (use independence)
Y

t=1
= 1 − [P (rt ≤ Q)]n (because of same distribution)

Consequently, let p be a small upper tail probability of rt and Q be

the corresponding quantile. That is,

P (rt ≤ Q) = 1 − p

From the above equation, we have

P (rn,i > Q) = 1 − (1 − p)n.

Therefore,

P (rn,i ≤ Q) = 1 − P (rn,i > Q) = (1 − p)n.

24
This means that Q is the (1 − p)n-th quantile of the generalized
extreme value distribution.

Takeaway: For a small probability p, compute (1 − p)n, where n

is the length of subperiod, then VaR can be obtained by finding the
(1 − p)n-th quantile of the extreme value distribution.

Credit Risk:
Reference: Credit Risk Measurement: New Approaches to Value
at Risk and Other Paradigms, 2nd Edition, by Anthony Saunders
and Linda Allen, Wiley, 2002.
Some techniques for credit risk measurement

1. Long-term credit rating (High to Low)

S&P Moody Fitch

AAA Aaa AAA
AA Aa AA
A A A
BBB Baa BBB
BB Ba BB
B B B
CCC Caa CCC
CC Ca CC
C C C
D D D

25
2. Credit quality over time (transition)

S&P One-year transition matrix

(Source: Standard & Poor’s, Feb. 1997)
Ini. Rating at year-end(%)
Rat. AAA AA A BBB BB B CCC Default
AAA 88.5 8.05 0.72 0.06 0.11 0.00 0.00 0.00
AA 0.76 88.3 7.47 0.56 0.05 0.13 0.02 0.00
A 0.08 2.32 87.6 5.02 0.65 0.22 0.01 0.05
BBB 0.03 0.29 5.54 82.5 4.68 1.02 0.11 0.17
BB 0.02 0.11 0.58 7.01 73.8 7.4 0.89 0.98
B 0.00 0.09 0.21 0.39 5.98 72.8 3.42 4.92
CCC 0.17 0.00 0.34 1.02 2.20 9.64 53.13 19.21

3. CreditMetrics (JP Morgan & other sponsors)

4. Altman Z score (Mainly is U.S.)

Z = 3.3(Earnings before Interest and Taxes [EBIT]/Totl Assets)

+ 1.0(Sales/Total Assets)
+ 0.6(Market Value of Equity/Book Value of Debt)
+ 1.4(Retained Earnings/Total Assets)
+ 1.2(Working Capital/Total Assets)

5. KMV Corporation’s credit risk model (now merged with Moody)

26
CreditMetrics: developed by J.P. Morgan and other sponsors in
1997.
Simply put, CreditMetrics addresses the question:
“How much will one lose on his loans and loan portfolios next year
for a given confidence level?”
From the assessment of market risk, the current market value and
its volatility of a financial position play an imporant role in VaR
calculation. Application of VaR methodology to nontrabable loans
encounters some immediate problems:

1. The current market value of the loan is not directly observable,

because most loans are not traded.

2. No time-series data available to estimate the volatility.

To overcome the difficulties, we make use of

1. Available data on a borrower’s credit rating

2. The probability that the rating will change over the next year
(the rating transition matrix)

3. Recovery rates on deafulted loans

4. Credit spreads and yields in the bond (or loan) market.

Example: Consider a five-year fixed-rate loan of $100 million made

at 6% annual interest, and the borrower is rated BBB.

27
Note: The numerical numbers used in this example are from Chap-
ter 6 of the reference book cited above.
Rating migration: One-year transition probabilities for BBB-
rated borrower
AAA AA A BBB BB B CCC Defaulty
0.02 0.33 5.95 86.93 5.30 1.17 0.12 0.18
Valuation Rating change (upgrades and downgrades) will affect
the required credit risk spreads or premiums on the loan’s remaining
cash flows and, hence, the implied market value of the loan.
Downgrade → credit spread premium rises → present value of the
loan should fall.
Upgrade has the opposite effect.
return to the example. (after one-year and a credit rating change)
6 6 6 106
P = 6+ + 2
+ 3
+ ,
1 + r1,1 + s1 (1 + r1,2 + s2) (1 + r1,3 + s3) (1 + r1,4 + s4)4
where r1,i are the risk-free rates on zero-coupon U.S. Treasury bonds
expected to exist one year into the future and si is the annual credit
spread on loans of a particular rating class of 1-year, 2-year, 3-year
and 4-year maturities (derived from observed spreads in the corporate
bond market over Treasuries).
One-year forward zero curves plus credit spreads by credit rating
category:

28
 Category Year 1 Year 2 Year 3 Year 4
AAA 3.60 4.17 4.73 5.12
AA 3.65 4.22 4.78 5.17
A 3.72 4.32 4.93 5.32
BBB 4.10 4.67 5.25 5.63
BB 5.55 6.02 6.78 7.27
B 6.05 7.02 8.03 8.52
CCC 15.05 15.02 14.03 13.52
Suppose that, during the first year, the borrower gets upgraded from
BBB to A. The present value of the loan is
6 6 6 106
P =6+ + + + = $108.66.
1.0372 (1.0432)2 (1.0493)3 (1.0532)4
Value of the loan at the end of Year 1, under different rating changes
(including first-year coupon):

Rating AAA AA A BBB BB B CCC Defaulty

value 109.37 109.19 108.66 107.55 102.02 98.10 83.64 51.13

Calculation of VaR

29
 New Loan Difference
Value Plus Probability of Value Probability
Year-end Probability Coupon Weighted from Weighted
Rating of State(%) (millions) Value($) Mean ($) Diff. Squared
AAA 0.02 109.37 0.02 2.28 0.0010
AA 0.33 109.19 0.36 2.10 0.0146
A 5.95 108.66 6.47 1.57 0.1474
BBB 86.93 107.55 93.49 0.46 0.1853
BB 5.30 102.02 5.41 (5.06) 1.3592
B 1.17 98.10 1.15 (8.99) 0.9446
CCC 0.12 83.64 1.10 (23.45) 0.6598
Default 0.18 51.13 0.09 (55.96) 5.6358
Form the table, the mean value of the loan is $107.09 (sum of the
4-th column). The variance of the value is 8.9477 (sum of the last
column).
√
Consequently, the standard deviation is σ = 8.9477 = 2.99.

If normal distribution of the loan value is used,

• 5% VaR: 1.65 × σ = $4.93

• 1% VaR: 2.33 × σ = $6.97

If actual distribution is used,

• 6.77% VaR: $107.09-102.02 = $5.07

30
 • 1.47% VaR: $107.09-98.10 = $8.99

• 1% VaR: $107.09-92.29 = $14.80.

The 1% number 92.29 is obtained by interpolation as (1.47%, 98.10)

and (0.3%, 83.64).

31