See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/284700328

Don't look back

Article · October 1998

CITATIONS READS

85 2,256

3 authors:

Giovanni Barone-Adesi Frederick Bourgoin

University of Lugano JP Morgan Asset Management
131 PUBLICATIONS   3,032 CITATIONS    5 PUBLICATIONS   95 CITATIONS   

SEE PROFILE SEE PROFILE

Kostas Giannopoulos
Neapolis University
30 PUBLICATIONS   731 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Financial Market Interdependencies View project

onparametric, conditional pricing of higher order multivariate contingent claims View project

All content following this page was uploaded by Kostas Giannopoulos on 10 December 2015.

The user has requested enhancement of the downloaded file.

 Market risk

Historical simulation may be a natural setting for scenario analysis,

but it must take account of current market conditions,
caution Giovanni Barone-Adesi, Frederick Bourgoin and Kostas Giannopoulos

aiu e-al-risk is becoming increasingly popubr as a man- and negative retu rns to have different impacts o n volatility ( known as the
3gcmenl and regulatory tool. BUI before Ihis acccp- Icvcr:.tge effect, see i3lack, 1976) . Past chilly p ortfoliO returns are divided
i:lIlce goes much furrher, wc need to :ISSCSS its rdi:lhility hy the Garch volatility estimated for the sa me date to obtain standardised
unt!l.:f fin:l1lcial 1l1:lrkd condit ions. Most VAR m odels n.!sidu:tl.... ThL'se are independen t and k k: nticdly dist ributed 0[1)) and arc
deal either wilh the non-normality of security returns therefore suitable for historical simulation .
or with their conditional heleroscedasticity, but not To ad just them to current market conditio ns, wc multiply a r:1ndomly
with ho th . We :II'C developing :1 modified hi.sloric:t1 simul:ition :Ippro:lch se lected stancbr<iised residua l hy the Garch forecast of tomorrow's volatil-
1hal allows for bol h effects. ity. In this w:JY,:t simu lated portfolio return for tolllorrow i. . uht:Jinco. Th i ~
I-lislorica l simul:lI ion rdies on a specific dislribution (usu:liiy uniform o r simulated return is used lO update the Garch forecast for the follOWing day,
normal) to se l~cl returns from the P:.lSl. These returns :lrc applied 10 cur- which is then multiplied by a newly sel ected, standardised residual to sim-
rent assel price!> (0 sim ulait:: their future returns. Once enough difTer(.!nt ula te the return for the second cia)' . Our recursive procedure is repc Jted
paths h:I VC been explored, il is po.ssible to determine a portfolio VAH with- until the VA R horizon (le, 10 d.IYs) is rc.lChcd, generating a sa mple path
out making ~lIb i tr.:lly assumptio ns ;Iho lH the distribution of ponfolio n..:lurns. of ponfolio voialil ilies :mcl returns. Wc n.:pe:lI our procedure to obtain a
This is especially u ~eful where there are abnonnally large ponfolio n..:turns. batch o f sample paths of portfolio returns. A confidence band for the cor-
I1 is well known th~Lt large returns cluster in time (see, for ex:.unple, responding portfolio values is built by taking th/": kernd (e mpirical) fre-
J"landelbrot, 1963, :lnd llbck, 1976). The resulting fluctuations in daily quency distribulion of values ~It each time. The lo wer 1% area identifies
volatility make lhe confidence levels or some VAR caicllbtions unrcliablt.! the worst GISt.' over the next 10 clays.
(Bou doukh et ai, 1995). Th is is the case with those that ignon.: c1ush.:ring, To illustrate our procedure , wc constructed a hypothetical portfoliO, di -
sudl :IS VAl{ Ille:tSlln:l1l<,::nts b:tsed on the standard variance-covariance ma- versi fied across all 13 national eqUity markets in o ur data sample. To form
trix and Mont e Carlo methods, w hich typically ignore current market con- o ur portfoliO, each equity market is weighted in proportion IQ its c<1pital-
ditions to produce flat volatility forecasts for future days. MoreoV<.::r, the use isation in the world index (MSC !) as at December 1995. The portfoliO
of the covariance matrix of security relurns or the choice of ;:1[) <Irbitrary weights are repolled in table A.
distribution in the Monte Carlo method usually destroys valuable infor- These weights arc held constant for the entire lO-year period and mul -
mation about the distribution of portfoliO returns. tiplied by the 13 loca l index returns. So the portfolio returns are calculat-
To make our historic:tI simul:.ition consistent with the clustering of large ed again backwards to reflect the current weightings. Si nce the aim of
returns, we model the volatility of our po rtfolio as an asymmetric Garch market risk is 10 quantify eventual portfo lio losses in a single cu rrency, all
(GenerJlised autoregressive conditional heteroscedasticity) process (Engle local portfoliO returns are measllred in dollars. The descriptive statistics, to-
& Ng, 1993) that generalises Ihe Garch model. This model allows pos itive gether with the Jarque-L3era (980) test fo r norma lity. arc shown in table 13,
where the p-va lue indicates the probabilit y that Ollr portfoliO retlJrns are
generated from a nonn ~tI d istribution.
Figure 1 shows the em pirical distribution of the portfolio's returns. The
Country reject ion of normality in tabl e A and the IXHtern of cl ustering visible in fig-
ure 1 leads us 10 modd our po rtfo lio returns, rt' as a Garch process with
asymmetries , with daily volatility, ht, give n by:
Denmark 0.004528
France 0.035857 (la)
Germany 0.039086 2
Hong Kong 0.0 17645 ht = U) + "(£t- 1 + y) + ph t - 1 (l b)
Italy 0.012709
The variance for small increments on the other end G ill be writt en as;
Japan 0.233527
Netherlands 0.022900 hr = crllt = O(lIt)
Singapore 0.006667 T he daily return in equation (la) is the sum of each expected value, J.1 ,
Spain 0.010254 plus a random residual, ~. Because of the sma ll, sta tist ica lly il~significant
Sweden 0.011 571 value of J.12, this term will be neglected in the calculatio n of daily volat ili-
Switzerland 0.033898 ties.! Equation (lb) defines the vola tilit y of £1' ht, as an asymmetric Garch
UK 0.096264 process. ht is the su m o f a consla nl, 00, plus two te rms reflecting the con-
US 0.407818 tributions o f the most recent "surprise", £t-l ' and the last period's volat ili-
ty, ht_I ' Finally, y allows for the asymmetric response of the innovation on
the volat ili(y and is statistically Significant.
Therefore, our portfoliO volatility is modelled 10 depend o n the most
recently observed portfoliO n.:l urns. The comb ination of asymmetric Garch
volatility and portfo liO historical returns offers us a fast and accu rate

1 In fact, for stock pn'ces, J12 is in the order of J..I.2tl[:

112 = C2M2 = 0(M2)

100· RISK· AUGUST 1998

 1. World capitalisation weighted portfolio 2. Portfolio stress analysis (standardised
returns: Jan 1985-Feb 1996 residuals): Jan 1986-Dec 1995

0,10 10

0,05 5

'@
"oc
~
os; 0
"c
c
ID
~ 'C
c 1"
~ -0.05 ~ -5
::;; c
'"
(jj
-j),10 -10

86 87 88 89 90 91 92 93 94 95

measure of the past, current and fururc vola tiJities of the current portfolio.
We do not need th e correlation m,llrix of security returns. Furthermo re, 3. Annualised volatility of the portfolio: Jan
our VAH method conlains fewer "unplc:1sant surprises", since Garch mod- 1986-Dec 1995
els allow for fa! tails on the unconditional d istribution of the data. l The ef-
fects of our choice become apparent if wc compare the returns in figure
I with those in figure 2, where they have been scaled by the ir dai ly volatil- 25
ity, so that:
rt 20
Zt =J; (2)

Clustering of returns is reduced by volatility sca ling, so the distribution

of returns now appears to be more uniform, making th e historical simula-
tion more appropriate. However, the large number of returns still exceed-
ing three sta ndard deviations suggests that our scaling does not make
relll rns normal. Our annualised portfolio volatil ity, show n in figure 3, var-
ied from 7% 10 21% over the lO-year period.
5
The sca led relllrns are the foundation of our simulation. To simulate
portfoliO retu rns over the next 10 days we select randomly 10 returns from
figure 2 llsing the "bOOtstr.:lp" methodology developed by Efron & Tibshi- O+---~-.--.---r--.---r--r--.--.-~
i
rani (993). We then construct iteratively the daily portfo lio volatility that 86 87 88 89 90 91 92 93 94 95 1
these returns imply accord ing to equations (1 a) and Ob) and lIse this volatil-
iry to resca le our returns. The resu lting returns thererore renect current
market conditions, rather than the m~l rke l condil ions associated w ith re- where i = 1 •. .. ,10 days and j = 1 •... •N, where N is Ihe number o f simu-
turns in figu re 1. In other words, we simulate future sta ndardised residual lation runs performed. The act u ~ll si mulated returns arc given by:
relurns as a random vector <3 of outcomes from a sta tionary distribution. , , ,.- (3a)
rl+i = tt+iVht +i
The possible outcomes ofl he stationary distribution are the historical resid-
uals, sland:mlised by the correspondi ng dail y vobli lity: where h; +i is :I (s imub[ ed) vobtilit y esti m:l1e obtained aoS:
.. = {£;j = el , e = {r"r2, ,, ,,rT} , (' )2,
ht+i = 00 + a tt+i-l + Y + Pht +i- l (3b)
2 For a Garch process with conditional normality, the excess of kurtosis of the un-
conditional distribution of the process is greater than three. See Bollersfev (1986) and:

2 Austin Friars, London EC2N 2HE Telephone Direct Line: 0171 4566915/456 69 16 General: 0171 456 6900
Fax: 0171 4566925 E-mail: search@michelangelo,CQ,uk
Please contact Ivan Collins or Alison Phillips
Michelangelo Associates
101 • RISK. AUGUST 1998
 Market risk

4. Normalised estimated distribution of returns 5. Estimated distribution of portfolio VAR In 10

in 10 days v. the normal density (10,000 days (10,000 simulations)
simulations)

0.45
- Smoothed density
0.40 ....... Normal density
0.35
0.30
0.25
....
0.20
0. 15
0.10
"- 0.05
0
-7.5 -5.0 -2.5 0 2.5 5.0 7.5
Standard deviations

options may be included by expressing their values in terms of assets meet~

ing our reqUirements, such as spot rates (for bonds) and underlying assets
~ is a random swnclarcli'sed residual estimated as in equat ion (2), but (for opt ions). A full re~eV:lIUalion procedure for these asselS ("an then be
rescaled to account for current market risk. In this W<IY, wc preserve the included at each step of our simulation (I3:1rone~Adesi, Giannopoulos &
time-series properties of the data. Vosper, 1997) . •
To obl:lin tht: distribution of our P011folio returns, wc replicate the above
procedure N == 10,000 times. The resulting (normalised) distribution is shown Giovanni Barone-Adesi is a professor at the University of Alberta,
in figure 4. The normal distribution is shown as a doned line for case of com- Canada and a visiting professor at City University Business School,
London. Frederick Bourgoin is an analyst at Millennium Global In-
parison. We may extend our procedure to multiple assets, preserving the COf-
vestments i n London. Ko stas Giannopoulos is a senior lecturer at th e
relations of asset returns by taking returns in the same day for each asset as University of We stminster, London . The suitability of our model for
input to our simubtion. Furthermore, unlike ordinal)' hi~torical simulmions, stress testing is being evaluated at th e London Clearing House (LCH).
it is possible to preselve autocorrelation and lagged cross correlation patterns We gratefully acknowledge LCH's continu ing support for our researc h
in the data by allowing past price changes (0 affect current returns.' e~maiJ: [email protected]
Not surprisingly, simu lated returns on our well~diversified portfolio are [email protected]
almost normal, except for sleeper peaking around zero and some dusler~
ing in the tails. The general share of the dislributidn supports the validity
of the usual measure of VA R for our portfolio. However, a closer exallli~
nation of our ~illlulation results show~ how even ou r well~diversified port~
folio m;!y depart from normality, There are, in fact, several occurrences of
very brge neg:l1ive returns, reaching :l maximum loss of 9.52%. Our em~
pi rical distribution implies losses of 3.380/0 and 2.24% at confidence levels
of 1% and 5% respectively.
The reason for this departure is the changes in portfolio volatility and
thus in pOl1folio VAR, as shO\vn in figure 5. The portfolio VA R over the
next 10 clays depends on the random returns seleClecl in each simulation
run. Its pattern is skewed to the right, shOWing how large returns tend to
cluster in time. These du~ters provide the base for reali~tic \Vorst~Glse sce~
nario :m:t1ysis consistt'nt with historical experience. To s(;'e the whole tlis·
tribution of \Vo rs t ~casc scenarios, wc need simply to repea t our simulation
and rcrord the worst~case scen:lrio of each run.
The worst~case sce n:lrio, as described in Boudoukh el aI, is defined as
the avel.lge of the outcomes in a given percentile. We h:tve extended their
approach by laking in to account the effect of time~varying volatility. O f
course, ollr method would produce more extreme dep:lrtures from nor~
nulity for less~diversificd ]1onfolios.
In conclusion. our simul ation methodology allows for fast ev:du:l1 ion
or VAR :tnd worst~C:lse scenario~ for brge portfol iOS. It t"kes into account
current market conditions and does not rely on knowledge of either the
correlation matrix of security returns or of the conditiona l distribution of
the underlying process. O ur methodology applies directly to asset returns
tl1:11 CII1 he !l1oddktl as condition:11 hett'roscccbstic processes . Bonds and
3 Only heteroscedasticity in this case. However, jf appropriate, autoregressive and
moving average returns can easily be inserted in equation (1 a) to maintain any other
properlles
10 3 · RISK · AUGUST 1998
View publication stats