Hilary Term Week 3 Assignment Solutions

lip.klimenka@economics.ox.ac.uk

February 11, 2015

Exercise 8.2

The gure below plots the daily returns on IBM from 1 January 2007 to 31 December 2007 (251 trading
days), along with 5% Value-at-Risk (VaR) forecasts from two models. The rst model (denoted HS) uses
`historical simulation' with a 250-day window of data. The second model uses a GARCH(1,1) model,
assuming that daily returns have a constant conditional mean, and are conditionally Normally distributed
(denoted `Normal-GARCH' in the gure).

Exercise 8.2

(1) Briey describe one other model for VaR forecasting, and
discuss its pros and cons relative to the `historical simulation'
model and the Normal-GARCH model.
CaViaR, or conditional autoregressive Value-at-Risk, is an
alternative to HS and Normal-GARCH that directly models the VaR
quantile of interest using a GARCH-like specication. In its
simplest form, the quantile is assumed to evolve according to

q,t = + I[rt1 <q,t1 ] + q,t1

where q,t is the quantile of the return distribution at time t ,
controls the response to a recent HIT and controls the
persistence. More exible forms can be introduced by allowing for
asymmetries across the sign of rt or for specifying an indirect
GARCH evolution.

Exercise 8.2

The strength of the CaViaR model is that it is narrowly focused on

estimating the relevant quantile, and so the parameters of the
model will be chosen to make it t the VaR as well as possible. As
a result, a CaViaR model can potentially correctly model the VaR
with relatively weak assumptions about the distribution of returns.
This is in contrast to a Normal-GARCH model which is a volatility
model  and optimized to t the conditional variance  and which
uses the quantile from a Normal irrespective of the actual quantiles
of standardized returns.
CaViaR diers from HS in that it uses conditional information. HS
directly estimates the quantile of the unconditional distribution of
returns but ignores conditioning information. Since it is well known
that volatility is clustered it should be expected that HS will have
periods with either far too many VaR violations or far too few.

Exercise 8.2

HittHS

= 0.0956 + ut

HittGARCH

= 0.0438 + ut

(0.0186)

(0.0129)

(i) How can we use the above regression output to test the
accuracy of the VaR forecasts from these two models?
These regressions are useful in testing for correct specication of a VaR
model using a restricted version of the classic Mincer-Zarnowitz
regression. The left-hand-side variable,

Hitt ,

measures when a VaR

exceedance has occurred, and so it should occur, on average,

time (where

of the

is the VaR target quantile, in this case 5%) . The sole

coecient measures the average occurrence rate of a

Hit

in the sample,

and so it can be used. For a correctly specied model, the null we would
be interested in testing is
where

is the intercept.

H 0 : 0 =

against an alternative

H1 : 0 6=

Exercise 8.2

(ii) What do the tests tell us?

We can construct a t -test for each which gives values of
0.0956 0.05
= 2.4516,
.0186

0.0438 0.05
= 0.4806
.0129

and so we would reject the rst and not the second since the
2-sided CV is 1.96.

Exercise 8.2

(b) Another set of regressions was also run (standard errors

are in parentheses below the parameter estimates):
HittHS
HittGARCH

HS
= 0.1018 0.0601Hitt
1 + ut
(0.0196)

(0.0634)

GARCH
= 0.0418 + 0.0491Hitt
+ ut
1
(0.0133)

(0.0634)

A joint test that the intercept is 0.05 and the slope

coecient is zero yielded a chi-squared statistic of 6.9679 for
the rst regression, and 0.8113 for the second regression.

Exercise 8.2

HittHS

HS
0.1018 0.0601Hitt
1 + ut

HittGARCH

GARCH
0.0418 + 0.0491Hitt
+ ut
1

(0.0196)

(0.0133)

(0.0634)

(0.0634)

(i) Why are these regressions potentially useful?

These regressions are useful in testing for correct specication of a VaR model
using a Generalized Mincer-Zarnowitz regression. The left-hand-side variable,
Hitt , measures when a VaR exceedance has occurred, and so it should occur
approximately of the time (where is the VaR target quantile, in this case
5%) and the occurrence of a Hit at time t should be unpredictable using
anything known at time t 1. In these regressions the intercepts are included
to test unconditional level and the lagged Hit is used to test if the occurrence
of a Hit at time t 1 is useful in predicting the occurrence of a Hit at time t .
For a correctly specied model, the null we would be interested in testing is
H0 : 0 = 1 = 0 against an alternative H1 : 0 6= 1 6= 0 where 0 is
the intercept. These regression are known as Dynamic Quantile regressions.

Exercise 8.2

(ii) What do the results tell us? (The 95% critical values for a
chi-squared variable with q degrees of freedom are given below:)
The null imposes two restrictions and so is distributed

22 ,

and the critical

value is 5.99. As a result the HS model is rejected while the

GARCH-based model is not.

q
1

95% critical value

3.84

5.99

7.81

9.49

11.07

10

18.31

25

37.65

249

286.81

250

287.88

251

288.96