Modelling of Long Term Risk

Contents
Modelling of Long Term Risk

A. Basel II
Roger Kaufmann
Department of Mathematics
B. Scaling of Risks
ETH Zurich
C. One-Year Risks
http://www.math.ethz.ch/kaufmann
D. Conclusions and Further Work
Risk Management lecture, ETH Zurich
June 10, 2004
c 2004 (R. Kaufmann, ETH Zurich)

1
A. Basel II Basel II (cont.)
Amendment to the Capital Accord to Incorporate Market Risks

Market risk: 10-day value-at-risk, 99%
(Basle Committee on Banking Supervision, 1996):
Standard: 1-day value-at-risk, 95%
In calculating the value-at-risk, a 99th percentile, one-tailed
confidence interval is to be used.
In calculating value-at-risk, an instantaneous price shock Insurance: 1-year value-at-risk, 99%

equivalent to a 10 day movement in prices is to be used. 1-year expected shortfall, 99%
Banks may use value-at-risk numbers calculated according to
shorter holding periods scaled up to ten days by the square root of
time.

2 c 2004 (R. Kaufmann, ETH Zurich)
3
Value-at-Risk and Expected Shortfall VaR in Visual Terms
Profit & Loss Distribution (P&L)
Primary risk measure: Value-at-Risk defined as Mean profit = 2.4
95% VaR = 1.6
0.25
1
VaRp(X) = FX (p) ,
0.20
i.e. the pth quantile of FX . (X denotes the profit, X the loss.)
probability density
0.15
Alternative risk measure: Expected shortfall defined as
0.10

ESp(X) = E X X < VaRp ;
0.05
i.e. the average loss when VaR is exceeded. Sp(X) gives 5% probability
information about frequency and size of large losses.
0.0
-10 -5 0 5 10

5
Losses and Profits B. Scaling

Loss Distribution
Mean loss = -2.4
Question 1: How to get a 10-day VaR (or 1-year VaR)?
0.25
95% VaR = 1.6
95% ES = 3.3

0.20
Solution in the praxis: scale the 1-day VaR by 10 (or 250).

probability density
0.10 0.15
Question 2: How good is scaling?

0.05
5% probability Model dependent!

0.0
-10 -5 0 5 10

7
Scaling under Normality Accounting for Trends
Under the assumption When adding a constant trend ,
i.i.d. i.i.d.
Xi N (0, 2), Xi N (, 2),
n-day log-returns are normally distributed as well: n-day log-returns are still normally distributed:
n
X n
X
Xi N (0, n 2). Xi N (n, n 2).
i=1 i=1
e2)-distributed profit X, VaRp(X) =

For a N (0, e xp, where xp Hence

denotes the p-Quantile of a standard normal distribution. Hence VaR(n) + n = n (VaR(1) + ),
i.e.
VaR(n) = n VaR(1).
VaR(n) = n VaR(1) (n n).

9
Autoregressive Models Scaling for AR(1) Models
For an autoregressive model of order 1, For an AR(1) model with normal innovations,
i.i.d. s
Xt = Xt1 + t, t N (0, 2), VaR(n) 1+ 1 n
= n 2 .
VaR(1) 1 1 2
1-day and n-day log-returns are normally distributed:

2
Xt N 0, For small values of , n VaR(1) is a good approximation of VaR(n).
1 2
and n
X 2 1 n
Xi N 0, 2
n 2 2
.
i=1
(1 ) 1

11
Non-Normal Innovations Non-Normal Innovations (cont.)
random walk, t8

Question: Is scaling with n still appropriate if innovations are
0.14
non-normal?
0.12
10dayVaR (99%); empirical
*
* **
0.10
i.i.d. **
Example: random walk, Xi t8 *
* * *
* *
*** ********************* ************* *
* * *
* **
* * *** ****** * ***
*
* ** ** *
* * *
*
*
**
*
** * *
*** ***************************************************************** ****** * *
0.08
* * *** * **** * *
**** *********************************************************************** **
(1) ********************************************
* ***** * * ************* **** ***** *** * *
************** ** **** * * ***
*
Based on 250 log-returns, how good is d 99% as an

10 VaR
** *********************************************************************************************** * *
************************************************************* ** **
*
0.06
** ****************** ** ***
* * ************ * *
**** * ***
estimate for the 10-day 99% VaR? *
0.04
(1)
d
(VaR99% denotes the one-day 99% VaR estimate.)
0.02
0.04 0.06 0.08 0.10 0.12 0.14
sqrt(10) * 1dayVaR (99%)
Scaling is still good, but other methods like random resampling

perform slightly better.
13
AR(1)-GARCH(1,1) Processes Scaling for AR(1)-GARCH(1,1) Processes
0.10

10-day, t4 innovations
scaled 1-day, t4 innovations
10-day, t8 innovations
A more complex process, often used for practical applications, is

scaled 1-day, t8 innovations
10-day, normal innovations
scaled 1-day, normal innovations

the GARCH(1,1) process ( = 0) and its generalization, the
0.09

AR(1)-GARCH(1,1) process:

Xt= Xt1 + tt,

0.08

t2= a0 + a(Xt1 Xt2)2 + b t1

2
,
t i.i.d., E[t] = 0, E[2t ] = 1. 0.0 0.05 0.10 0.15 0.20
(typical parameters: = 0.04, a0 = 3 106, a = 0.05, b = 0.92) Goodness of fit of the scaling rule, depending on different values of
(x axis) for different distributions of the innovations t.
For typical parameters ( = 0.04, t t8), the fit is almost perfect.
15
GARCH(1,1) vs. Random Walk GARCH(1,1) vs. Random Walk (cont.)
A GARCH(1,1) process If the initial values of the processes (Xa,t) and (X0,t) coincide, then
Xa,t = a,tt, E[(Xa,t X0,t)2] fct(parameters),

2 2 2
a,t = a0 + a Xa,t1 + b a,t1 ,
and
t i.i.d., E[t] = 0, E[2t ] = 1, n+h
X n+h
X
E[( Xa,t X0,t)2] fct(parameters).
(where a is typically close to 0) can be approximated by a process t=n+1 t=n+1
with variance
2 2 These inequalities can be used to get bounds for (conditional and
0,t = a0 + b 0,t1
unconditional) value-at-risk of GARCH(1,1) processes. Analogously,
or value-at-risk estimates for AR(1)-GARCH(1,1) processes can be
2 2
0,t = a0 + (a + b) 0,t1 . obtained by approximating them with AR(1) processes.
17
Stochastic Volatility Model with Jumps Stochastic Volatility Model: Volatility and Returns
An alternative to autoregressive types of models are stochastic

volatility models:
1.8
1.6
Xt = a t Z t + b J t t ,
1.4

t = t1 ec Yt ,
1.2
i.i.d.
t, Zt, Yt N (0, 1),
1.0
Jt Bernoulli()
0.8
(typical parameters: 0 50 100 150 200 250 0 50 100 150 200 250
= 0.01, a = 0.01, b = 0.05, c = 0.05, = 0.98)

19
Scaling in the Stochastic Volatility Model C. One-Year Risks
0.20
10day VaR
scaled 1day VaR
0.16
Problems when modelling yearly data:
Non-stationarity of data sets.

0.12
Lack of yearly returns.

0.08
0.0 0.02 0.04 0.06 0.08 0.10 Properties of yearly data are different from those of daily data.
Goodness of fit of the scaling rule, depending on different values of
(x axis).
The scaled 1-day VaR underestimates the 10-day VaR for small
values of . For > 0.04, this changes to an overestimation.
21
How to Estimate Yearly Risks Models
Fix a horizon h < 1 year, for which data can be modelled. Random Walks
Use a scaling rule for the gap between h and 1 year. Autoregressive Processes
GARCH(1,1) Processes
scaling rule
Heavy-tailed Distributions
suitable model
today h days 1 year

23
Random Walk Autoregressive Processes
Financial log-data (st)thN can be modelled as a randow walk For an AR(p) model with trend and normal innovations,
process with constant trend and normal innovations:
p
X
i.i.d. st = ai stih + t for t hN,
st = sth + Xt, Xt N (, 2) for t hN. i=1
(t N (0 + 1 t, 2), independent)

The square-root-of-time rule (accounting for the trend) can be used
to scale h-day risks to 1-year risks. the 1-year value-at-risk and expected shortfall can be calculated as a
function of the parameters 1, and ai, and the current and past
values of (st).

25
Generalized Autoregressive Conditional Heavy-tailed Distributions

Heteroskedastic Processes
h-day log-returns (Xt)thN are said to have a heavy-tailed
Assuming a GARCH(1,1) process with Student-t distributed distribution, if
innovations for h-day log-returns,
P [Xt < x] = xL(x) as x ,
Xt = + t t for t hN,
t2 = 0 + 1(Xth )2 + 1th
2
, where R+ and L is a slowly varying function,
i.e. limx L(sx)
L(x) = 1 for all s > 0.
i.i.d.
where t t , E[t] = 0, E[2t ] = 1,
Also in this case, 1-year VaR and ES can be estimated based on the
1-year log-returns follow a so-called weak GARCH(1,1) process. The parameter and on the observed data.
corresponding VaR and ES can be calculated as a function of the
above parameters and the current and past values of (Xt).
27
Backtesting Conclusions for 1-Year Forecasts
The suitability of these models for estimating one-year financial risks

The random walk model performs in general better than the other
can be assessed by comparing estimated value-at-risk and expected
models under investigation. It provides satisfactory results across
shortfall with observed return data for
all classes of data and for both confidence levels investigated (95%,
99%). However, like all the other models under investigation, the
stock indices,
risk estimates for single stocks are not as good as those for foreign
foreign exchange rates, exchange rates, stock indices, and 10-year bonds.
10-year government bonds, The optimal calibration horizon is about one month. Based on
these data, the square-root-of-time rule (accounting for trends) can
single stocks. be applied for estimating one-year risks.

29
Confidence Intervals for a Random Walk D. Conclusions

0.6
0.5
The square-root-of-time scaling rule performs very well to scale

risks from a short horizon (1 day) to a longer one (10 days, 1 year).
0.4
* * *
The reasons for this good performance are non-trivial. Each
0.3
* *
* * *
* *
situation has to be investigated separately. The square-root-of-

0.2
h = 1 day h = 1 week h = 1 month h = 3 months time rule should not be applied before checking its appropriateness.
h = 1 year
0.1
In the limit, as 1, scaling a short-term VaR to a long-term

Point estimates and 95% confidence intervals for one-year risk using the square-root-of-time rule is for most situations not
99% expected shortfall and 99% value-at-risk (percentage loss) for a appropriate any more.
simulated random walk with normal innovations.
31
Further Work Bibliography
[Brummelhuis and Kaufmann,

2004a] Brummelhuis, R. and Kauf-
An interesting subject for further research is to find the limits,
mann, R. (2004a). 10-rule for GARCH(1,1) and AR(1)-
where the square-root-of-time rule fails. For example changing
GARCH(1,1) processes. Working Paper.
one single parameter in a model can have a strong effect on the
appropriateness of this scaling rule. [Brummelhuis and Kaufmann, 2004b] Brummelhuis, R. and Kauf-
mann, R. (2004b). GARCH(1,1) and AR(1)-GARCH(1,1)
Linked to this topic is the model-dependent question, why the
processes: perturbation estimates for value-at-risk. Working Paper.
square-root-of-time rule performs well (or not so well) in a certain
situation. [Embrechts et al., 2004] Embrechts, P., Kaufmann, R., and Patie, P.
(2004). Strategic long-term financial risks: Single risk factors.
An interesting generalisation of this work would be the investigation
To appear in: A special issue of Computational Optimization and
of multivariate models.
Applications.
33

[Kaufmann, 2004] Kaufmann, R. (2004). 10-rule for stochastic
volatility models. Working Paper.
[Kaufmann and Patie, 2003] Kaufmann, R. and Patie, P. (2003).

Strategic long-term financial risks: The one-dimensional case.
RiskLab Report. Available at http://www.risklab.ch/Papers.html
#SLTFR.

34

Modelling of Long Term Risk

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Modelling of Long Term Risk

Uploaded by

Copyright:

Available Formats

Contents

Modelling of Long Term Risk

c 2004 (R. Kaufmann, ETH Zurich)

A. Basel II Basel II (cont.)

Amendment to the Capital Accord to Incorporate Market Risks

In calculating value-at-risk, an instantaneous price shock Insurance: 1-year value-at-risk, 99%

c 2004 (R. Kaufmann, ETH Zurich)

information about frequency and size of large losses.

c 2004 (R. Kaufmann, ETH Zurich)

Losses and Profits B. Scaling

95% VaR = 1.6

Solution in the praxis: scale the 1-day VaR by 10 (or 250).

Question 2: How good is scaling?

5% probability Model dependent!

c 2004 (R. Kaufmann, ETH Zurich)

e2)-distributed profit X, VaRp(X) =

c 2004 (R. Kaufmann, ETH Zurich)

Autoregressive Models Scaling for AR(1) Models

c 2004 (R. Kaufmann, ETH Zurich)

Based on 250 log-returns, how good is d 99% as an

Scaling is still good, but other methods like random resampling

AR(1)-GARCH(1,1) Processes Scaling for AR(1)-GARCH(1,1) Processes

A more complex process, often used for practical applications, is

Xt= Xt1 + tt,

t2= a0 + a(Xt1 Xt2)2 + b t1

t i.i.d., E[t] = 0, E[2t ] = 1. 0.0 0.05 0.10 0.15 0.20

Xa,t = a,tt, E[(Xa,t X0,t)2] fct(parameters),

An alternative to autoregressive types of models are stochastic

= 0.01, a = 0.01, b = 0.05, c = 0.05, = 0.98)

c 2004 (R. Kaufmann, ETH Zurich)

Non-stationarity of data sets.

Lack of yearly returns.

How to Estimate Yearly Risks Models

today h days 1 year

c 2004 (R. Kaufmann, ETH Zurich)

(t N (0 + 1 t, 2), independent)

c 2004 (R. Kaufmann, ETH Zurich)

Generalized Autoregressive Conditional Heavy-tailed Distributions

The suitability of these models for estimating one-year financial risks

c 2004 (R. Kaufmann, ETH Zurich)

Confidence Intervals for a Random Walk D. Conclusions

The square-root-of-time scaling rule performs very well to scale

situation has to be investigated separately. The square-root-of-

In the limit, as 1, scaling a short-term VaR to a long-term

[Brummelhuis and Kaufmann,

[Kaufmann and Patie, 2003] Kaufmann, R. and Patie, P. (2003).

c 2004 (R. Kaufmann, ETH Zurich)

You might also like

Xt= Xt1 + tt,

t i.i.d., E[t] = 0, E[2t ] = 1. 0.0 0.05 0.10 0.15 0.20

Xa,t = a,tt, E[(Xa,t X0,t)2] fct(parameters),

(t N (0 + 1 t, 2), independent)