Vous êtes sur la page 1sur 4

45

AENORM 56 August 2007


GARCH model s i n opt i on t r adi ng
For many aspect s of nanci al anal y si s uncer t ai nt y i s of maj or i mpor t ance. The condi t i onal
v ar i ance of nanci al t i me ser i es i s used f or cal cul at i ng measur es of r i sk of hol di ng an asset ,
t he const r uct i on of hedge por t f ol i os and opt i on pr i ci ng. For some nanci al model s, such as
t he CAPM ( Capi t al Asset Pr i ci ng Model ) and Bl ack - Schol es model , r et ur ns ar e assumed t o be
i ndependent ( nor mal l y ) di st r i but ed and var i ances i n r et ur ns ar e assumed t o be const ant . Thi s
assumpt i on about r et ur ns i s t y pi cal l y not sat i sed i n nanci al t i meser i es. One of t he most
i mpor t ant char act er i st i cs of r et ur ns of nanci al asset s or i ndi ces i s t hat t hei r mean appear s
t o be const ant w hi l e t hei r var i ance changes over t i me. Fur t her mor e, r et ur ns ex hi bi t vol at i l i t y
cl ust er i ng: l ar ge changes i n r et ur ns ar e l i k el y t o be f ol l ow ed by f ur t her l ar ge changes. These
char act er i st i cs can be seen f r om Fi gur e 1, t hat show s t he dai l y l og r et ur n of t he Eur o St ox x 50
I ndex , a w ei ght ed i ndex of 50 Eur opean st ock s.
Kar l i j n Jut t mann
st udied Economet rics at t he Universit y of Amst erdam. This
year she nished her Mast er in Financial Economet rics. She
did an int ernship at t he Asset Management depart ment of
I nsinger de Beaufort where she wrot e her t hesis GARCH
Models in Opt ion Trading. Last year she worked as an int ern
at Bain & Company in Chicago. Current ly Karlij n is working at
AlpI nvest Part ners as a Port folio and Risk Analyst .
-8
-6
-4
-2
0
2
4
6
8
1992 1994 1996 1998 2000 2002 2004 2006
R
Figure 1: Daily log returns of Euro Stoxx,
r
t
=100*ln(S
t
/S
t-1
)
One of t he rst economist s t o model t he above
ment ioned charact erist ics of ret urns was Robert
Engle when he int roduced t he Aut oregressive
Condit ional Het eroskedast icit y ( ARCH) model in
1982. Bollerslev ( 1986) ext ended t his model t o
t he Generalised ARCH model ( GARCH) . Since
t heir int roduct ion a great number of papers ap-
peared applying t hese models and ext ensions
t o it . The biggest st rengt h of t he GARCH mod-
els is t heir abilit y t o describe t he charact erist ics
above in t he variance of nancial ret urn dat a.
There is ext ensive evidence t hat GARCH mod-
els are highly signicant in- sample, but t here
is less evidence t hey also will provide good
forecast s of fut ure ret urn volat ilit y. The volat il-
it y forecast abilit y of t he GARCH models is not
st raight forward t o t est , because act ual volat ilit y
is not observed.
I n t his paper I will t est t he forecast abilit y of t he
GARCH model by using t he volat ilit y forecast s
in opt ion t rading. I will const ruct opt ion t rading
st rat egies based on a comparison of volat ilit y
forecast s given by t he GARCH model and t he
volat ilit ies t hat are implied by t he Black- Scholes
model and t he market opt ion prices. This implied
volat ilit y can be regarded as t he market s pre-
dict ion of fut ure volat ilit y. I f t he GARCH model
is capable of giving a bet t er volat ilit y forecast ,
t his forecast can be used t o const ruct a prot -
able st rat egy of buying and selling opt ions.
For t he est imat ion of t he GARCH model I will
use daily ret urn dat a of t he Euro St oxx 50 I ndex
bet ween 1992 and 2006. The opt ions used for
t rading are European put opt ions wit h t he Euro
St oxx index as t he underlying asset .
The remainder of t his art icle is organized as
follows. The next sect ion describes t he GARCH
model in some more det ail. Aft er t hat I will
discuss t he implied volat ilit ies from t he Black-
Scholes model and in t he last sect ions I will
explain how t he opt ion t rading st rat egies are
const ruct ed and what t he result s from t hese
st rat egies are.
GARCH Model s
As not ed before, GARCH models are well suit ed
t o describe most charact erist ics of ret urn dat a.
This sect ion describes t he GARCH model.
Dene t he log ret urn on a nancial asset as
Economet r i cs
46 AENORM 56 August 2007
r
t
= ln( S
t
/ S
t- 1
) , where S is t he price of t he asset ,
in t his case t he index. We can model t his log
ret urn as:
c

= O +
1
|
t t t t
r E r ( 1)
wit h

c o = , (0, 1)
t t t t
z z IID ( 2)
and
o c

= O = O

2 2
1 1
| |
t t t t t
E Var r ( 3)
where

O
1 t
is t he informat ion available up t o
t ime t- 1. Equat ion 2 implies t hat
t
is dist rib-
ut ed wit h mean zero and a t ime varying vari-
ance o
2
t
.
I n t he st andard GARCH( 1, 1) model t he variance
at t ime t is modelled as a linear funct ion of t he
square of t he past shock and t he condit ional
variance at t ime t- 1.
o e oc |o e o |

= + + > >
2 2 2
1 1
, 0, , 0.
t t t
( 4)

I t is possible t o include more lagged squared
shocks or lagged variances in t he model
but applied work frequent ly shows t hat t he
GARCH( 1, 1) model is suit able for represent ing
t he maj orit y of nancial t ime series. Under cer-
t ain rest rict ions t his model exhibit s mean re-
version t o an uncondit ional variance. I n most
st udies t here is a high persist ence in volat ilit y,
which means t he reversion t o t he uncondit ional
variance goes very slowly.
Several ext ensions t o t his model exist , which I
include in my invest igat ion. One of t hem is t he
TARCH ( Threshold GARCH) model, in which a
negat ive shock has a larger impact on fut ure
volat ilit y t han a posit ive shock of t he same
size.
o e oc c c |o

= + + < +

2 2 2 2 2
1 1 1 1
0
t t t t t
I ( 5)
A second ext ension I consider is t he Component
TARCH or CTARCH model. This model allows
mean reversion t o a level m
t
t hat changes over
t ime:
o o c
c c | o
e p e m c o



= + +
< +
= + +
2 2
1 1
2 2
1 1 1 1 1
2 2
1 1 1
( )
( ) [ 0] ( )
( ) ( )
t t t t
t t t t t
t t t t
m m
m I m
m m
( 6)
I n t his model o
2
t
converges t o t he t ime varying
long run volat ilit y m
t
, while m
t
converges t o
wit h powers of . is usually bet ween 0. 99 and
1 so t hat m
t
approaches very slowly.
I est imat e t hese models wit h t he Euro St oxx
ret urn dat a.
1
All of t he models above are able t o
describe t he charact erist ics of t he ret urn dat a
in- sample; t hey can remove t he het eroskedas-
t icit y from t he dat a. The asymmet ric or TARCH
model is preferred t o t he st andard GARCH
model, based on t he value of it s log likelihood
and t he signicance of t he paramet ers in t he
model.
Based on t he paramet er est imat es of t hese
models I make volat ilit y forecast s of t he fut ure
volat ilit y during t he lifet ime of t he opt ions used
for t rading.
I mpl i ed Vol at i l i t y
Anot her way t o const ruct a forecast of fut ure
volat ilit y during t he lifet ime of an opt ion is by
means of t he Black- Scholes model and opt ion
market prices. The Black- Scholes model is t he
most famous model t o price European opt ions.
One of t he det erminant s of opt ion prices is vol-
at ilit y, which is assumed t o be a const ant . By
put t ing all ot her det erminant s and market op-
t ion prices in t he model we can calculat e t he
volat ilit y implied by t he model. This can be seen
as a volat ilit y forecast over t he lifet ime of t he
opt ion which I will compare wit h t he volat ilit y
forecast s of t he previous sect ion.

Figure 2: Implied volatility and GARCH volatility
forecasts over the remaining lifetime of an option

Figure 2 shows t he implied volat ilit y t oget her
wit h t he volat ilit y forecast s given by t he differ-
ent GARCH models. On every t rading day be-
t ween 2001 and 2006 I t ake an at - t he- money
opt ion wit h t he lowest t ime t o mat urit y. For t his
opt ion I calculat e bot h volat ilit y forecast s. The
gure shows t hat t he t hree different GARCH
forecast s are quit e similar. They also move
close t oget her wit h t he implied volat ilit y, but
t he lat t er is on average slight ly higher t han t he
GARCH forecast s. This is especially t he case
during t he years 2003 and 2004. Based on t he
comparison of t hese t wo volat ilit ies I will con-
1
All models where est imat ed wit h t he assumpt ion of st udent - t dist ribut ed error t erms. This led t o bet t er result s t han
under t he assumpt ion of normally dist ribut ed error t erms.
Economet r i cs
48 AENORM 56 August 2007
st ruct an opt ion t rading st rat egy for every day
in t he sample.
Tr adi ng St r at egi es
For each t rading day bet ween 2001 and 2006
implied volat ilit ies and GARCH forecast s are con-
st ruct ed in t he previous sect ions. Furt hermore,
I const ruct a delt a neut ral port folio on each day
in t he sample. This port folio consist s of one op-
t ion and an amount of delt a
2
invest ed in t he
index. Figure 3 shows how t he value of a delt a
neut ral port folio changes wit h a change in t he
price of t he index.
Based on t he comparison of t he t wo volat ilit y
forecast s I creat e buy and sell signals for t he
port folios on every t rading day. I f t he GARCH
forecast exceeds t he implied volat ilit y of t he
opt ion a long posit ion in t he port folio is t ak-
en, which is held for one day. This means we
buy t he port folio and sell it one day lat er. I f t he
GARCH forecast is correct and t he change over
t his day in t he index is larger t han expect ed by
t he market , t his posit ion will be wort h more at
t he end of t he holding period. Figure 3a shows
t his. Selling t he posit ion one day lat er will t hen
lead t o a posit ive ret urn. I f t he GARCH forecast
will be lower t han t he implied volat ilit y a sell
signal is given and a short posit ion in t he port -
folio is t aken. I f t he forecast is correct and t he
change in t he index will be relat ively small, t his
also will lead t o a posit ive ret urn as can be seen
in Figure 3b.
Each t rading day t he above t rading st rat egy will
generat e a prot / loss of t he difference bet ween
t he port folio values ( ) at t imes t and t+ 1. I
calculat e t he daily ret urns ( R
t
) as t he prot as a
percent age of t he underlying index price ( S) .

+
H H
=
1
( )
t t
t
t
R
S
for a long posit ion and
+
H H
=
1
( )
t t
t
t
R
S
for a short posit ion ( 7)
I n t his way I const ruct ret urn series of st rat e-
gies based on t he GARCH, TARCH and CTARCH
models. I t is somet imes argued t hat opt ions
are consist ent ly overpriced which means t hat
implied volat ilit y will be always higher t han re-
alised volat ilit y ( Engle and Rosenberg, 1994) .
I n t his case a st rat egy of always selling op-
t ions will lead t o signicant prot s. As can be
seen from Figure 2 implied volat ilit y exceeds
t he GARCH forecast s on most days and a sell
signal is given. I f t he GARCH st rat egies gen-
erat e posit ive ret urns t his could be t he result
of overpriced opt ions. Therefore I also consider
t he ret urns of a st rat egy of selling t he opt ion
every day t o invest igat e if t he GARCH model
st rat egy has some addit ional value over an al-
ways sell st rat egy. I also const ruct a second,
lt ered st rat egy where a buy or sell signal is
given only if t he difference bet ween t he GARCH
forecast and t he implied volat ilit y exceeds a
cert ain value.
Resul t s
Table 1 shows t he ret urns from t rading put
opt ions according t o t he different st rat egies.
These ret urns are only relat ive because of t he
way t hey are calculat ed. I f a t rade in t wo op-
t ions was made inst ead of one, ret urns would
be t wice as large. However, t his way of calcu-
lat ing ret urns allows us t o compare t he result s
of t he different t rading st rat egies considered
here.
Ret urns from all st rat egies t urn out t o be signif-
icant ly posit ive for t he full sample, 2001- 2006.
2
The delt a of an opt ion is dened as t he rat e of change of t he opt ion price wit h respect t o t he change in t he price of
t he underlying asset and calculat ed as t he rst derivat ive of t he Black- Scholes pricing formula t o t he asset price.
Economet r i cs
Figure 3: The change in a delta neutral portfolio given a change in de underlying asset for a) a long position in
the portfolio and b) a short position in the portfolio
49
AENORM 56 August 2007
Considering t he t hree different sub samples,
we see t hat posit ive ret urns are only generat ed
in t he sample cont aining t he years 2003 and
2004. This result implies a superior forecast
abilit y of t he GARCH models over t he volat il-
it y forecast implied by t he opt ion prices in t he
period 2003- 2004.
However, Figure 2 shows t hat in t his sub sample
implied volat ilit y exceeds t he GARCH forecast
on most of t he days. I n 90% of t he days a sell
signal is given. Moreover, t he GARCH st rat egy
ret urns do not exceed t he ret urns of t he always
sell st rat egy. This suggest s t hat t he posit ive re-
t urns in 2003- 3004 are t he result of overpriced
opt ions in t his sub sample. Anot her way t o show
t his is t o consider t he ret urns aft er a buy and
sell signal separat ely for t he GARCH st rat egies.
The average ret urns generat ed aft er selling t he
port folio are signicant ly posit ive. The ret urns
aft er a buy signal are not . This means t hat t he
GARCH models are unable t o det ect t he days in
t he sample where implied volat ilit y is t oo low
and opt ions are underpriced.
The models perform well in t he 2003- 2004 sub
sample. Figures 1 and 2 show t hat in t his period
a t ransit ion t akes place from a high t o a low
volat ilit y period. The opt ion prices in t he mar-
ket are not able t o adj ust fast enough t o t his
change in volat ilit y. The GARCH models on t he
ot her hand make a fast er adj ust ment t o a lower
volat ilit y and t herefore sell signals are given on
most of t he days in t his sample, leading t o posi-
t ive ret urns.
Higher ret urns are generat ed by t he lt ered
st rat egies, where we only buy or sell opt ions
if t he difference bet ween t he GARCH forecast
and implied volat ilit y is sufcient ly large. The
lt ered TARCH model gives t he best result s and
out performs t he always sell model in t he period
from 2001- 2004. I n t he last sub sample, 2005-
2006, hardly any t rades are made and most of
t he ret urns in t his period are consequent ly zero.
Therefore, a good way t o use t he lt ered st rat -
egy is t o make an invest ment in anot her asset ,
for example t he risk free asset , on days no buy
or sell signal is given by t he model. Alt hough
t he lt ered st rat egies out perform t he unlt ered
and always sell st rat egies, t his model is st ill un-
able t o det ect t he days on which implied volat il-
it y is t oo low.
Concl usi ons
I n t he lit erat ure about GARCH models, t here
is ext ensive evidence t hat t he model performs
well in describing in- sample ret urn dat a. There
is less evidence t hat t he model is also capable
of making accurat e forecast s about fut ure vola-
t ilit y. This paper showed again t hat t he GARCH
models performed well in- sample. I t est ed t heir
forecast abilit y by using t he forecast s t o creat e
prot able opt ion t rading st rat egies. The mod-
els only performed well in t he years 2003- 2004
where a t ransit ion t ook place from a period of
high volat ilit y t o a period of low volat ilit y. The
opt ions in t his sample were overpriced and t he
GARCH models were able t o det ect t his.
The lt ered st rat egy performed bet t er and was
capable of generat ing posit ive ret urns in t he
period 2001- 2002 as well. However, none of t he
models were able t o det ect t he days on which
implied volat ilit y was t oo low and opt ions were
underpriced.
GARCH TARCH CTARCH al l w ay s
sel l
Full
sample
0. 013
( 0. 199)
0. 009
( 0. 199)
0. 013
( 0. 199)
0. 012
( 0. 199)
Sample
2001-
2002
- 0. 003
( 0. 265)
- 0. 003
( 0. 265)
0. 003
( 0. 265)
- 0. 004
( 0. 265)
Sample
2003-
2004
0. 030
( 0. 193)
0. 026
( 0. 193)
0. 025
( 0. 194)
0. 030
( 0. 193)
Sample
2005-
2006
0. 009
( 0. 108)
0. 003
( 0. 108)
0. 009
( 0. 108)
0. 006
( 0. 108)
Table 1: Average daily returns from trading
delta hedged put options (st.dev.)
Economet r i cs

Vous aimerez peut-être aussi