Journal of Empirical Finance 1 (1993) 83-106.

North-Holland

A long memory property of stock market

returns and a new model*

Zhuanxin Ding, Clive W.J. Granger and Robert F. Engle

Unrwrsily of Calfwxia, San Diego, CA, USA

Received June 1992. Final version accepted February 1993

Abstract

A long memory property of stock market returns is investigated in this paper. It is found that
not only there is substantially more correlation between absolute returns than returns them-
selves, but the power transformation of the absolute return lrfl also has quite high autocorrel-
ation for long lags. It is possible to characterize lrfld to be long memory and this property is
strongest when d is around 1. This result appears to argue against ARCH type specifications
based upon squared returns. But our Monte-Carlo study shows that both ARCH type models
based on squared returns and those based on absolute return can produce this property. A new
general class of models is proposed which allows the power 6 of the heteroskedasticity equation to
be estimated from the data.

1. Introduction

If Y, is the return from a speculative asset such as a bond or stock, this paper
considers the temporal properties of the functions 1I,(~ for positive values of d. It
is well known that the returns themselves contain little serial correlation, in
agreement with the efficient market theory. However, Taylor (1986) found that
/t-,1 has significant positive serial correlation over long lags. This property is
examined on long daily stock market price series. It is possible to characterize
(rtld to be long-memory, with quite high autocorrelations for long lags. It is also
found, as an empirical fact, that this property is strongest for d= 1 or near
1 compared to both smaller and larger positive values of d. This result appears
to argue against ARCH type specifications based upon squared returns. The
paper examines whether various classes of models are consistent with this
observation. A new general class of models is then proposed which allows the
power 6 of the heteroskedasticity equation to be estimated from the data.

Correspondrnce to: Clive Granger, University of California, San Diego, CA 92093-0508, USA.
*We thank Jurg Barlocher, Xiaohong Chen, Takeo Hoshi, Bruce Lehman, Victor Ng, and Ross
Starr for helpful comments and discussions. We are also grateful to the editor (Richard T. Baillie)
and two anonymous referees for their constructive comments. The second and third authors would
like to thank NSF for financial support.

0927-5398/93/$06.00 s$c 1993-Elsevier Science Publishers B.V. All rights reserved

84 Z. Ding et al., A long memory properly ofstock market returns

The remainder of this paper is organized as follows: In section 2, we give

a brief description of the data we use. In section 3 we carry out the autocorrela-
tion and cross-correlation analysis. The special pattern of the autocorrelogram
and crosscorrelogram of the stock returns is exploited and presented. Section 4
investigates the effect of temporal aggregation on the autocorrelation structure
and examines the short sample autocorrelation property of stock returns.
Section 5 presents a Monte-Carlo study of various financial models. Based on
this, we propose a new general class of models in section 6. Section 7 concludes
the analysis.

2. The data

The data set we will analyze in this paper is the Standard & Poor 500
(hereafter S&P 500) stock market daily closing price index. There are alto-
gether 17055 observations from Jan 3, 1928 to Aug 30, 1991. Denote pt as the
price index for S&P 500 at time t (t= 0, . . . , 17055). Define

I, = lnp, - lnp, _ 1 (1)

as the compounded return for S&P 500 price index at time t (t = 1, . . . , 17054).
Table 2.1 gives the summary statistics for rt. We can see from table 2.1 that the
kurtosis for rl of 25.42 is higher than that of a normal distribution which is 3. The
kurtosis and studentized range statistics (which is the range divided by standard
deviation) show the characteristic fat-tailed behavior compared with a normal
distribution. The Jarque-Bera normality test statistic is far beyond the critical
value which suggests that r1 is far from a normal distribution.
Figs. 2.1,2.2 and 2.3 give the plots of pt, Y,and Ir,l. We can see from the figures
the long run movement of daily pt, rl, lrtl over the past 62 years. There is an
upward trend for pr but Y, is rather stable around mean ~=0.00018. From the
series Ir,], we can clearly see the observation of Mandelbrot (1963) and Fama
(1965) that large absolute returns are more likely than small absolute returns to

Table 2.1
Summary statistics of T,.

sample studentized normality

data size mean std skewness kurtosis min max range test

r1 17054 0.00018 0.0115 -0.487 25.42 -0.228 0.154 33 357788

We are indebted to William Schwert for providing us the data.

 Z. Ding et al.. A long memory property of stork market returns 85

Fig. 2.1. Standard & Poor 500 daily price index 01/03/28-08/30/91.

Fig. 2.2. Standard & Poor 500 daily returns 01/04/28~8/30/91.

Fig. 2.3. Standard & Poor 500 daily absolute returns 01/04/28%08/30/91.

be followed by a large absolute return. The market volatility is changing over

time which suggests a suitable model for the data should have a time varying
volatility structure as suggested by the ARCH model. During the Great De-
pression of 1929 and early 193Os, volatilities are much higher than any other
period. There is a sudden drop in prices on Black Mondays stock market crash
of 1987, but unlike the Great Depression, the high market volatility did not last
very long. Otherwise, the market is relatively stable.

3. Autocorrelation analysis of the return series

It is now well established that the stock market returns themselves contain
little serial correlation [Fama (1970), Taylor (1986)] which is in agreement with
the efficient market theory. But this empirical fact does not necessarily imply that
86 2. Ding et al.. A long memory properry of stockmarket returns

returns are independently identically distributed as many theoretical financial

models assume. It is possible that the series is serially uncorrelated but is
dependent. The stock market data is especially so since if the market is efficient,
a stocks price should change with the arrival of information. If information
comes in bunches, the distribution of the next return will depend on previous
returns although they may not be correlated.
Taylor (1986) studied the correlations of the transformed returns for 40 series
and concluded that the returns process is characterized by substantially more
correlation between absolute or squared returns than there is between the
returns themselves. Kariya et al. (1990) obtained a similar result when studying
Japanese stock prices. Extending this line we will examine the autocorrelation of
rl and Ir,ld for positive d in this section, where r1 is the S&P 500 stock return.
Table 3.1 gives the sample autocorrelations of r,, 1r,I and r: for lags 1 to 5 and
10, 20, 40, 70, 100. We plot the autocorrelogram of rr, lr,l and rf from lag 1 to
lag 100 in fig. 3.1. The dotted lines show f 1.96/a which is the 95% confi-
dence interval for the estimated sample autocorrelations if the process rr is
independently and identically distributed (hereafter i.i.d.). In our cask T= 17054
so + 1.96/fi=O.O15. It is proved [Bartlett (1946)] that if r, is a i.i.d
process then the sample autocorrelation pr is approximately N(0, l/T). In
fig. 3.1, about one quarter of the sample autocorrelations within lag 100 are
outside the 95% confidence interval for a i.i.d process. The first lag autocorrela-
tion is 0.063 which is significantly positive. Many other researchers [see Fama
(1976), Taylor (1986), Hamao et al. (1990)] also found that most stock market

Table 3.1
Autocorrelations of rr.

data lag 1 2 3 4 5 10 20 40 70 100

rr 0.063 -0.039 -0.004 0.031 0.022 0.018 0.017 0.000 0.000 0.004
IT11 0.318 0.323 0.322 0.296 0.303 0.247 0.237 0.200 0.174 0.162
r, 0.218 0.234 0.173 0.140 0.193 0.107 0.083 0.059 0.058 0.045
 Z. Ding et al., A long memory property of stockmarket returns 87

return series have a very small positive first order autocorrelation. The small
positive first order autocorrelation suggests that the Y, do have some memory
although it is very short and there is a portion of stock market returns that is
predictable although it might be a very small one. So the efficient market or
random walk hypothesis does not hold strictly. Alternatively, this could be from
non-synchronous measurement of prices. The second lag autocorrelation
(= - 0.039) is significantly negative which supports the so called mean-rever-
sion behaviour of stock market returns. This suggests that the S&P 500 stock
market return series is not a realization of an i.i.d process.
Furthermore, if r, is an i.i.d process, then any transformation of rr is also an
i.i.d process, so will be 1I, / and Y: The standard error of the sample autocorrela-
tion of 1Y,1will be 1/fi= 0.015 if rt has finite variance, the same standard error
is applicable for the sample autocorrelation of r, providing the rt also have finite
kurtosis. But from fig. 3.1, it is seen that not only the sample autocorrelations of
(rt( and r: are all outside the 95% confidence interval but also they are all
positive over long lags. Further, the sample autocorrelations for absolute
returns are greater than the sample autocorrelations for squared returns at every
lag up to at least 100 lags. It is clear that the S&P 500 stock market return
process is not an i.i.d process.
Based on the finding above, we further examined the sample autocorrelations
of the transformed absolute S&P 500 returns 1rt Id for various positive d, Table 3.2
gives corr(Ir,Jd, jr,+, I) for d=0.125, 0.25, 0.50, 0.75, 1, 1.25, 1.5, 1.75, 2, 3 at
lags 1 to 5 and 10, 20,40, 70, 100. Figs. 3.2, 3.3 show the autocorrelogram of lrtld
from lag 1 to 100 for d= 1,0.50, 0.25,0.125 in fig. 3.2 and d= 1, 1.25, 1.5, 1.75, 2
in fig. 3.3. From table 3.2 and figs. 3.2, 3.3 it is seen that the conclusion obtained
above remains valid. All the power transformations of the absolute return have
significant positive autocorrelations at least up to lag 100 which supports the
claim that stock market returns have long-term memory. The autocorrelations
decrease fast in the first month and then decrease very slowly. The most

Table 3.2
Autocorrelations of lrf1

d lag 1 2 3 4 5 10 20 40 70 100

0.125 0.110 0.108 0.102 0.098 0.121 0.100 0.100 0.095 0.065 0.089
0.25 0.186 0.181 0.182 0.176 0.193 0.164 0.164 0.148 0.120 0.131
0.5 0.257 0.255 0.263 0.251 0.259 0.222 0.221 0.192 0.166 0.165
0.75 0.297 0.299 0.305 0.286 0.29 1 0.246 0.241 0.207 0.180 0.173
0.318 0.323 0.322 0.296 0.303 0.247 0.237 0.200 0.174 0.162
1.25 0.319 0.326 0.312 0.280 0.295 0.227 0.211 0.174 0.153 0.138
1.5 0.300 0.309 0.278 0.242 0.270 0.192 0.170 0.136 0.122 0.106
1.75 0.264 0.276 0.228 0.192 0.234 0.149 0.125 0.095 0.088 0.073
2 0.218 0.234 0.173 0.140 0.193 0.107 0.083 0.059 0.058 0.045
3 0.066 0.088 0.036 0.025 0.072 0.019 0.009 0.004 0.006 0.003
88 2. Ding et al., A long memory property of stock market returns

Fig. 3.2. d- 1, 0.5, 0.25, 0.125 from high to low.

Fig. 3.3. d- 1, 1.25, 1.50, 1.75, 2 from high to low.

Fig. 3.4. Autocorrelation of (r (**d at lag 1

interesting finding from the autocorrelogram is that 1I, Id has the largest autocor-
relation at least up to lag 100 when d = 1 or is near 1. The autocorrelation gets
smaller almost monotonically when d goes away from 1.
To illustrate this more clearly, we calculate the sample autocorrelations p,(d)
as a function of d,d>O, for r=l,2,5, 10 and taking d=0.125,0.130,. .,
1.745, 1.750, 2, 2.25, . . . , 4.75, 5. Figs. 3.4, 3.5, 3.6 and 3.7 give the plots of cal-
culated p,(d) at T= 1,2, 5, 10. It is seen clearly from these figures that the
autocorrelation p,(d) is a smooth function of d. There is a saddle point 2 be-
tween 2 and 3 such that when d cd, p,(d) is a concave function and when
d >& p,(d) is a convex function of d. There is a unique point d* around 1 such
that p,(d) reaches its maximum at this point, p,(d*) >p,(d) for d #d*.
 Z. Ding et al., A long memory property of stock market returns 89

Fig. 3.5. Autocorrelation of Irl **d at lag 2.

Fig. 3.6. Autocorrelation of Irl**d at lag 5.

0 1 2 3 4 5

Fig. 3.7. Autocorrelation of Irl**d at lag 10.

In fact, Ir,ld has positive autocorrelations over a much longer lags than 100.
Table 3.3 shows the lags (r*) at which the first negative autocorrelation of lrtld
occurs for various d. It can be seen from the table that in most cases, lrtjd has
positive autocorrelations over more than 2500 lags. Since there are about 250
working days every year, the empirical finding suggests that lrrld has positive
autocorrelations for over 10 years!
90 2. Ding et al., A long memory propertyof stockmarketreturns

Table 3.3
Lags at which first negative autocorrelations of lrrld occurs.

d 0.125 0.25 0.5 0.75 1 I .25 1.5 1.75 2 3

T* 2028 2534 2704 2705 2705 2705 2705 2685 2598 520

:
0 500 loo0 1500 2oca

Fig. 3.8. Autocorrelation of /rl up to lag 2500.

We pick 1r, 1as a typical transform of the return series here and plot its sample
autocorrelations up to lag 2500 in fig. 3.8. The dotted lines are 95% confidence
interval for the estimated sample autocorrelation of an i.i.d process as before. It
is striking that all the sample autocorrelations are not only positive but also stay
outside the confidence interval. Different models have been tried to approximate
this sample autocorrelation curve, including: (1) pr an exponentially decreasing
function of z( pI = a/?) (which is similar to the autocorrelation function of
a ARMA model); (2) pr the same as the autocorrelation function of a fraction-
ally integrated process [see Granger and Joyeux (1980)]

r(1-m r(T+B)
&= I-(P) T(T+l -B)

=-u-b? (T+B-l)...BUb)
(2)
W) (T-m...(l-P)n-P)
(T+P-l)...fl
=(T-j?)...(l-/I)

=p
z
Jr+P-l).
(T-P)
and (3) pT a polynomially decreasing function of z(p, = a/?) which is approxim-
ately the same as (2) when t is large. It is found, compared to the real data, that
the fitted autocorrelation using method (1) decreases too slowly at the beginning
 Z. Ding et al.. A long memory property qfstock market returns 91

and then too fast at the end while by using methods (2) and (3) the opposite
result is found.
The final preferred model is a combination of these methods. A theoretical
autocorrelation function is specified as follows:

which can easily be transformed to a linear model

lOgp,=lOga+~,lOgp,-,+TlOgP,-B,lOgT. (4)

Let cr*=loga, pT=/r, bT=logpz, and /?3= --p3, then

lOgp,=a*+~~lOgp,~,+~~T+~flOgT. (5)

Ordinary Least Squares gives as estimates:

log&= -0.049+0.784lOgp,_,-0.195X 10-4T-o.05710gT, (6)

(- 3.9) (62.9) (- 5.9) (-9.1)

R= = 0.92, D-W= 2.65.

The t-statistics inside parentheses show that all the parameters are significant.
After transferring the above equation back to autocorrelations one gets:

p*=o.893p,o~,84(0.999955)/z~~~~. (7)

Fig. 3.9 plots the fitted autocorrelations (dotted line) and the sample autocorre-
lations themselves. It is seen that the theoretical model fits the actual sample
autocorrelations quite well.
Similar studies were also carried out for the New York Stock Exchange daily
price index and the German daily stock market price index (DAX) over a shorter
sample period (1962-1989 for NYSE, 198&1991 for DAX); we get similar

Fig. 3.9. Autocorrelation of It-1(solid line) and its fitted value (dotted line).
92 Z. Ding et al.. A long memory property qf stock market returns

autocorrelation structures for transformed returns. Furthermore, we did the

cross-correlation analysis of transformed S&P 500 and New York Stock Ex-
change daily returns series and also found the cross-correlation is the biggest
when d= 1 and that it also has long memory. This suggests there may be
volatility co-persistence for these two stock market index prices (see Bollerslev
and Engle 1989). Our conjecture is that this property will exist in most financial
series.

4. Sensitivity of autocorrelation structure

We now further investigate the effect of temporal aggregation on the autocor-

relation structure. Table 3.4 gives the autocorrelations of 1I;, 5Id, where yt;,5 is the
5 day temporal average of Y,. i.e.

r;,
5 =-

:( r,,+l+r,f+2+ ..+rtf+5),

where t=l,2,..., 3410 and t=5(t-1). It can be seen that the temporal
aggregation does not change the long memory property of the absolute return
series. pr( 1I;, 5I) still reaches a unique maximum when d is around 1 or 1.25 for
different lags r. Compared with the original daily series, the first order autocor-
relation for 1ft;,5 Id is much bigger than the second one. Although the temporally
aggregated return series here is not exactly the same as a weekly return series, we
expect a similar result will hold for the weekly data.
It should also be noted from fig. 2.2 that the volatility structure differs
considerably between the pre-war and the post-war period. The pre-war period
(1928-1945) is much more volatile than the post-war period (19461986). It will
be interesting to look at the memory structure for these two periods. Table 3.5

Table 3.4
Autocorrelations of 1f,, 51.

d lag 1 2 3 4 5 10 20 40 70 100

0.125 0.145 0.109 0.148 0.149 0.136 0.105 0.129 0.077 0.072 0.041
0.25 0.187 0.155 0.184 0.184 0.169 0.137 0.158 0.102 0.095 0.052
0.5 0.247 0.213 0.229 0.227 0.204 0.180 0.188 0.136 0.126 0.065
0.75 0.296 0.255 0.26 1 0.255 0.223 0.212 0.203 0.161 0.149 0.074
1 0.332 0.279 0.279 0.267 0.227 0.233 0.205 0.175 0.163 0.079
1.25 0.352 0.286 0.282 0.263 0.217 0.243 0.197 0.178 0.168 0.080
1.5 0.356 0.277 0.271 0.245 0.196 0.242 0.180 0.173 0.164 0.076
1.75 0.349 0.255 0.250 0.217 0.169 0.23 1 0.160 0.160 0.153 0.069
2 0.332 0.227 0.223 0.186 0.140 0.214 0.138 0.144 0.138 0.061
3 0.231 0.109 0.115 0.075 0.048 0.124 0.068 0.079 0.073 0.026
 Z. Ding PI al., A long memory property of stockmarket returns 93

Table 3.5
Autocorrelations of )r,ld 1928-1945.

d lag 1 2 3 4 5 10 20 40 70 100

0.125 0.114 0.135 0.126 0.117 0.138 0.131 0.122 0.118 0.067 0.115
0.25 0.201 0.227 0.23 1 0.204 0.215 0.200 0.197 0.183 0.128 0.158
0.5 0.273 0.298 0.31 I 0.275 0.276 0.245 0.245 0.216 0.169 0.172
0.75 0.300 0.323 0.332 0.296 0.294 0.25 1 0.248 0.212 0. I72 0.162
1 0.310 0.329 0.329 0.296 0.293 0.241 0.232 0.192 0.159 0.141
I .25 0.310 0.323 0.310 0.280 0.28 1 0.223 0.205 0.163 0.138 0.1 16
1.5 0.302 0.310 0.283 0.256 0.260 0.199 0.173 0.130 0.114 0.090
1.75 0.289 0.292 0.25 1 0.226 0.236 0.175 0.141 0.099 0.09 1 0.067
2 0.273 0.272 0.218 0.196 0.211 0.151 0.111 0.072 0.070 0.047
3 0.201 0.194 0.114 0.098 0.128 0.076 0.034 0.012 0.020 0.007

Table 3.6
Autocorrelations of lrfld 19461986.

d lag 1 2 3 4 5 10 20 40 70 100

0.125 0.089 0.062 0.054 0.057 0.086 0.047 0.054 0.051 0.04 1 0.038
0.25 0.129 0.095 0.086 0.102 0.126 0.082 0.082 0.068 0.058 0.053
0.5 0.162 0.128 0.121 0.141 0.158 0.111 0.106 0.082 0.068 0.066
0.75 0.181 0.151 0.143 0.164 0.175 0.126 0.119 0.088 0.067 0.068
1 0.191 0.167 0.157 0.180 0.182 0.133 0.123 0.089 0.062 0.064
1.25 0.194 0.178 0.163 0.191 0.180 0.134 0.120 0.084 0.053 0.056
1.5 0.189 0.182 0.160 0.198 0.170 0.129 0.110 0.074 0.042 0.046
1.75 0.178 0.179 0.150 0.200 0.154 0.119 0.095 0.06 1 0.03 I 0.036
2 0.163 0.170 0.135 0.199 0.133 0.105 0.078 0.047 0.02 1 0.027
3 0.099 0.104 0.066 0.173 0.056 0.047 0.023 0.010 0.002 0.005

shows the autocorrelations of 11-1~for the pre-war period (1928-1945). It is seen

that the magnitude of the autocorrelation for lrrld is about the same as those in
table 3.2. (~~1has the largest autocorrelation for the first two lags and then this
property becomes strongest for Jr,l.75 or lr110.5.
Table 3.6 gives the autocorrelations of (rtld for the post-war period
(1946-1986). It is clear from the table that during this less volatile period the
market has both a smaller and a shorter memory in the sense that the autocorre-
lations are smaller and decrease faster. The autocorrelations are only about two
thirds as big as those of the pre-war period.
Comparing table 3.2, 3.5 and 3.6 we can probably say that the long
memory property that was found in the whole sample period can be mainly
attributed to the pre-war period. The market has a strong and long memory
of big events like the great depression in 1929 and early 1930s when volatility
was very high.
94 Z. Ding et al., A long memory property of stock market return.~

5. Monte-Carlo study of various financial time series models

The empirical findings of section 3 and 4 have strong implications for the
modeling of financial time series. Taylor (1986) showed neither day-of-the-
week effects nor a linear, correlated process can provide satisfactory explanation
of the significant correlations among absolute return series, where a linear
correlated process can be represented as

Y,=i+ f CtiC?-i. (9)

i=O

where r and G(iare constants with c(,,= 1, e, is a zero-mean i.i.d process. Taylor
concludes that any reasonable model must be a non-linear one. Furthermore,
the special autocorrelation pattern of lrrld found in section 3 implies that any
theoretical model should be able to capture this before the model can be
considered to be adequate.
It should be noted that a process can have zero autocorrelations but have
autocorrelations of squares greater than for moduli. For example, consider the
following nonlinear model:

rt = IstIet, (10)

where e,-N(0, l), E(s,)=E(q,)=E(r,)=O, IaI< 1, e, and qt are stochastically

independent, s, is independent of v,+~ for T > 0, s,, s,_, are jointly normally
distributed with variance 1, hence we have var( qr) = 1 - ~1~and qt - N(0, 1 -a*).
The conditional variance of r, when s, is known is s:, i.e. var( r, Is,) = s:. For this
model corr(r,, r,_,)=O but by using numerical integration it is found that with
Ial<

2 2 rX2
corr(Ir,I, lr,_,J)=- EIs,s,_,+txsf-- <corr(r:,rf_,)=q. (11)
71( 7c1

It is thus seen that the results of table 3.2 do not necessarily occur.
One possible explanation for the large positive autocorrelation between lrtl
and IY~+~I or lr,ld and IY,+,(~ is the heteroskedasticity of the data, i.e. the variance
or conditional variance is changing over time. One family of nonlinear time
series models that is able to capture some aspects of the time varying volatility
structure is Engles ARCH (AutoRegressive Conditional Heteroskedasticity)
model [Engle (1982)]. In its original setting, the ARCH model is defined as
a data generating process for a random variable which has a conditional normal
 Z. Ding et al., A long mrmory property ofsstock market returns 95

distribution with conditional variance a linear function of lagged squared

residuals. More formally, the ARCH(p) model is defined as follows:
Y,= /l + EI

h=w,, e,-N(O, I), (12)

P

S:=cCo+ C Sli&r2_i.
i=l

It is easily shown that Y,is not autocorrelated with each other but 1rr Id is. Hence
the distribution of rr is dependent on rt_;, i >O. Since its introduction by Engle
(1982), the ARCH model has been widely used to model time-varying volatility
and the persistence of shocks to volatility. Much work has also been done both
theoretically and empirically. Many modifications and extensions of the original
ARCH model have also appeared in the literature.
For example, in order to capture the long memory property of the conditional
variance process, Bollerslev (1986) introduced the GARCH(p, q) model, which
defines the conditional variance equation as follows:

(13)

Taylor (1986) modeled the conditional standard deviation function instead of

conditional variance. Schwert (1989) following the argument of Davidian and
Carroll (1987) modeled the conditional standard deviation as a linear function
of lagged absolute residuals. The Taylor/Schwert GARCH(p, q) model defines
the conditional standard deviation equation as follows:

(14)
i=l j=l

One may, at first glance, think that it would be better to use Taylor/Schwert
model than Bollerslevs GARCH since the model is expressed in terms of
absolute returns rather than squared returns. But this conclusion is not neces-
sarily true when the model is a nonlinear one. In fact, our Monte-Carlo study
shows both Bollerslevs GARCH and Taylor/Schwerts model with appropriate
parameters can produce the special correlation patterns found in section 3.
Both models were estimated for S&P 500 returns and the following results
were obtained:
(I) GARCH
r,=0.000438+0.144~,_,+e,,
(7.2) (18.4)
s:=0.0000008+0.09
l~:_~ +0.906s;. (15)
(12.5) (50.7) (43.4)
log likelihood: 56822.
96 2. Ding et ul., A long memory property qfstock market returns

(2) TaylorlSchwert
r,=0.0004+0.139&,_r+&,,
(7.0) (19.6)

s,=0.000096+0.104(~,_,~+0.913s,. (16)
(12.6) (67) (517)

log likelihood: 56776.

The first order moving average term is in the mean equations of both models
to account for the positive first order autocorrelation for the return series. We
can see all the parameters are very significant in the above models. The
normality test statistic of the standardized residuals for both models are far
beyond the critical value of a normal distribution as assumed by both models.
This is not surprising since there are definitely other factors affecting the
volatility. Nevertheless, the log-likelihood value for Bollerslevs GARCH is
significantly larger than that of Taylor/Schwert model.
Based on the estimation results, some simulations have performed using the
parameters estimated above assuming e, - IID N(0, 1). Our purpose is to check
whether theoretical ARCH models can generate the same type of autocorrela-
tions as stock market return data. Obviously if the theoretical model does not
exhibit the same pattern of autocorrelations as stock market return data, then it
follows that the theoretical model is misspecified for these data. A total of 18054
observations was generated and the first 1000 were discarded in order to be less
affected by the initial value of s0 which was set to be the unconditional standard
deviation of the S&P 500 returns. Figs. 4.1,4.2,4.3 and 4.4,4.5,4.6 plot the
simulated autocorrelogram of the data generated by the two models. It can be
seen that the special autocorrelation pattern does exists here. For both models,
Jrld has the largest autocorrelation when d= 1, and the autocorrelation gets
smaller when d goes away from 1. It is interesting that Bolleslevs GARCH
model can produce this result even though the conditional variance is a linear
function of squared returns. For Bollerslevs GARCH model, the autocorrela-
tion between I r, ) and )rt + r 1is very close to that between lrf(1.25 and )rt+,(1.25. But

Fig. 4.1. Bollerslevs GARCH model. Autocorrelation of jrl, r**2, r from high to low
 97
2. Ding et al., A long memory propurtq of stock market returns

n a0 .

Fig. 4.2. Bollerslevs GARCH model. d- 1, 0.5, 0.25, 0.125 from high to low.

Fig. 4.3. Bollerslevs GARCH model. d- 1, 1.25, 1.50, 1.75, 2 from high to low.

Fig. 4.4. Taylor/Schwert model. Autocorrelation of Irl. r**2, r from high to low

Fig. 4.5. Taylor/Schwert model. d- I, 0.5, 0.25, 0.125 from high to low

for the Taylor/Schwert model, the autocorrelation between Jrl) and Irr+r 1after
lag 40 is close to that between IY,[,~ and Irt+r l.5. One major difference between
autocorrelograms of the two simulated data series and the real data is that the
autocorrelations of the real data decreases rapidly in the first month and then
decrease very slowly over a long period, but the autocorrelations of the two
simulated data decrease almost constantly over time.
98 2. Ding et al.. A long memory property qf stock market returns

Fig. 4.6. Taylor/Schwert model. d- 1. 1.25, 1.50, 1.75, 2 from high to low.

6. A new model - Asymmetric power ARCH

The Monte-Carlo study shows that the ARCH model generally captures the
special pattern of autocorrelation existing in many stock market returns data.
Both Bolleslevs GARCH and Taylor/Schwerts GARCH in absolute value
model can produce this property. It seems there is no obvious reason why one
should assume the conditional variance is a linear function of lagged squared
returns (residuals) as in Bollerslevs GARCH, or the conditional standard
deviation a linear function of lagged absolute returns (residuals) as in
Taylor/Schwert model. Fortunately, a more general class of model is available
which includes Bolleslevs GARCH, Taylor/Schwert and five other models in
the literature as special cases. The general structure is as follows:

St=9ZO+i CCi(lE-i/-Yi&t_i)+ 5
j= 1
BjSf-i, where (17)
i=l

a,>O, 630,

c(i>O, i=l, . , p,

-1 <yi< 1, i=l, . , p,

fij>O, j= 1, . . ,fJ

The model imposes a BoxxCox power transformation of the conditional stan-

dard deviation process and the asymmetric absolute residuals. By using this
transformation we can linearize otherwise nonlinear models. The functional
form for conditional standard deviation is familiar to economists as the constant
elasticity of substitution (CES) production function of Arrow et al. (1961).
The asymmetric response of volatility to positive and negative shocks is well
known in the finance literature as the leverage effect of the stock market returns
[Black (1976)], which says that stock returns are negatively correlated with
changes in return volatility ~ i.e. volatility tends to rise in response to bad news
(excess returns lower than expected) and to fall in response to good news
 Z. Dinget al., A long mrmor~ propertyof stock market returns 99

(excess returns higher than expected) [Nelson (1991)]. Empirical studies by

Nelson (199 l), Glosten, Jaganathan and Runkle (1989) and Engle and Ng (1992)
show it is crucial to include the asymmetric term in financial time series models
[for a detailed discussion, see Engle and Ng (1992)].
This generalized version of ARCH model includes seven other models (see
appendix A) as special cases. We will call this model Asymmetric Power ARCH
model and denote it as A-PARCH.
If we assume the distribution of rt is conditionally normal, then the condition
for existence of Es: and El &,I6 is (see appendix B):

R-l

& izlcCi{(l +yi)6+( l_Yi)')2 * 1.(F)+j=li flj< l. (18)

If this condition is satisfied, then when 632 we have E, convariance

stationary. But 6 32 is a sufficient condition for E, to be covariance
stationary.
The new model is estimated for S&P 500 return series by the maximum
likelihood method using the BerndttHall-Hall-Hausman algorithm. The estim-
ated model is as follows:

r,=0.00021+0.145~,~, +q,
(3.2) (19.0)

St1~43=O.OOOO14+O.O83(~s,-~~-O.373~,_,)~43+O.92Os~~43, (19)
(4.5) (32.4) (-20.7) (474)(33.7)

log likelihood: 56974.

The estimated 6 is 1.43 which is significantly different from 1 (Taylor/

Schwert model) or 2 (Bollerslev GARCH). The t-statistic for the asym-
metric term is 32.4 which is very significant implying the leverage effect does
exist in S&P 500 returns. By using the log-likelihood values estimated, a
nested test can easily be constructed against either Bollerslevs GARCH or
Taylor/Schwert model. Let lo be the log-likelihood value under the null hypo-
thesis that the true model is Bollerslevs GARCH and 1 be the log-likelihood
value under the alternative that the true model is A-PARCH, then 2(1-lo)
should have a x2 distribution with 2 degrees of freedom when the null hypo-
thesis is true. But in our example 2( 1- 1,,) = 2(56974- 56822) = 304 which is
far beyond the critical value at any reasonable level. Hence we can reject that
the data is generated by Bollerslevs GARCH model. The same procedure
is applicable to Taylor/Schwert model and we can also reject it.
100 Z. Ding et al., A long memory property of stock market returns

6. Conclusion

In this paper, a long-memory property of the stock market returns series is

investigated. We found not only there is substantially more correlation between
absolute returns than returns themselves, but the power transformation of the
absolute return Jr,ld also has quite high autocorrelation for long lags. Further-
more, for fixed lags, the function p,(d)=corr(Ir,Id, JT~+~I~)has a unique max-
imum point when d is around 1. This result appears to argue against ARCH type
specifications based upon squared returns. But our Monte-Carlo study shows
both ARCH type of model based upon squared return and those based upon
absolute return can produce this property. The ARCH specification based upon
the linear relationship among absolute returns is neither necessary nor sufficient
to have such a property. Finally, we propose a new general class of ARCH
models which we call Asymmetric Power ARCH model and denote A-PARCH.
The new model encompasses seven other models in the literature. We estimate
S&P 500 returns by the new model and the estimated power 6 for the condi-
tional heteroskedasticity function is 1.43 which is significantly different from
1 (Taylor/Schwert model) or 2 (Bollerslevs GARCH).

Appendix A

We now show that the new model includes the following seven ARCH models
as a special case.

(1) Engles ARCH(p) model [see Engle (1982)], just let 6 = 2 and yi =O,
i=l,..., p, fij=O, j= 1, . , q in the new model.
(2) Bollerslevs GARCH(p, q) model (see Bollerslev 1986), let 6 =2 and yi=O,
i= 1, . . . ) p.
(3) Taylor/Schwerts GARCH in standard deviation model let 6 = 1 and yi = 0,
i=l 9 . . 2
(4) GJR mode![see Glosten et al. (1989)], let 6=2.

When 6 = 2, we have when 0 d yi < 1

S:=Z,+ ~ ri(lE,_il-Yi&,_i)2+ ~ BjS:-j

i=l j=l

=cCo+ ~ cci(1-_i)2&:_i+ ~ j3jS:_j

i=l j=l

+ ~ ~~{(l+~~)Z~(l~~~)Z}S~&~-~
i=l
 Z. Ding et al., A long memory property qfstock market returns 101

=a,+ f ri(l-Yi)&rZ_i+ f BjS,2-j+ f_ 4CXiYiS;Ef-i, where

i=t j=l i=l

1 if El-i<0
s; =
0 otherwise.
If we further define
aT=cci( 1 -_Yi)2)
YT=4cciyi,
then we have

S$Cr,+ i aT&:_i+ i fljSf_j+ f Y:S;E,Z_i

i=l j=l i=l

which is exactly the GJR model.

When - 1 <y,<O we have

S:=&,+ f Cri(l+yi)&:-i+ i PjSf-j- f 4aiyiSi+&,Z-i, where

i=l j= 1 i=l

1 if E*_i>O
s+=
0 otherwise,
define
cC~=Ui(l +yi)2,
y? = -4(q]Q,

we have

Sf=OZQ+ i ai*&;_i+ i BjSf-j+ i vTSi+E,Z_i

i= 1 j=l i=l

which allows positive shocks to have a stronger effect on volatility.

(5) Zakoians TARCH model (see Zakoian 1991), let 6 = 1 and flj=O,
j=l,. * , q. We have

St=c(o+ i Ui( IE,-il-yiE,-i)

i= 1

where
i=l i=l

+
&,-i if E,-i>O
&,-i= and
0 otherwise,
Ef-i=&,_i-E,-i.
+
102 2. Ding et al., A long memory property cfslock market returm

So by defining

Cc+=ai( 1 -yi),

yi- =Ei( 1 +/i)r

we have

which is the exact TARCH form. If we further let Pj #O, j = 1, . . , q then we get
a more general class of TARCH models.

(6) Higgins and Beras NARCH model [see Higgins and Bera (1990)], let
yi=o, i= 1, . 3 p and bj=O, j= 1, . . . , 4.
Our model becomes

S~=Cr,+ ~ C(il&t~ilb, i.e.

i=l

(sf)6/2=Eo+ J$ Cli(E:-i)SZ.
i=l

Define

We have exactly Higgins and Beras NARCH.

(7) Geweke (1986) and Pantula (1986)s log-ARCH model. The log-ARCH
model is the limiting case of our model when d-+0.
Since

SP=cCi)+ 2 ai(lEt-il-_Yi&t-i)6+ i fljS:-j,

i=l j=l

decompose a0 as:

c(O= l- i C4iE(le,_il-)rie,_i)a- i pj
i i=l j=l

=uXd,

hence Es: =ws.Then we have

s;-1
--= 1- i aiE(le,_il-yie,_i)6- i pi v
6 i i=l j= 1 I
 Z. Ding el al.. A long mrmory property ofstockmarket rc~ums 103

+ i a,( lEt-il-l/iEl-i)d- l sr-j- 1

I +i/ljF
i=l 6 j=l

~(le,_i)-yie,_i)6-l
-ii1c(i 6

when 640 the above equation becomes

log&= l- i zilimE(Ie,-i/-yie,_j)- i bj lOgo

i i=1 6-O j=1 I

+ i ailOg((E,-i(-yjE,-i)+ i /?jlOgSf_j
i=l j=l

-i$l C(ilOgE(Ie,-iI-yie,-i)

=ZXlOgW- $J UilOg&+ i ~ilOg(l&-il-Yiet-i)

i=l i=l

+ i PjlOgst-j,
j=l

where

LX;= {I- i ccilimE(le,_il-yie,~i)s- i pi)

i=l 6-O j=l

= 1-~ ~i-~ Bj )
i i=l j=l I

since lim,, 0 E( Ie, _ i I- yie,_ i)6 = 1. This is a generalized version of Geweke/Pan-

tulamodel.Ifwefurtherletyi=O,i=l,. . . ,p,and/Ij=O,j=I,. , q,thenwe
get the exact Geweke/Pantula model.

Appendix B. Conditions for the existence of Esq\and El ~~1

If we assume the distribution is conditional normal, then the condition for

existence of Es: of the nevy model is

$J giE((e,-i)-yie,-i)6+ i /?j<l, where

i=l j=l

+cC x2
1
E(lC?-il-Yiet-i)6=p (lxl-~ix)c2 dx
4% s --s(i
104 Z. Ding it al., A long tnemor~ properly ofstockmarket return

So the condition becomes

h i$lXi[(l +/i)+(l
-(r :_1+i
-yi)6]2611 j=l
pj<l. (Bl)

Since
EI&,I~=Ele,(~Es:

So the condition for the existence of Ejatld is the same as that of Est. The proof
of the above result is almost identical to the proof of theorem 1 in Bollerslev
(1986).
When condition (Bl) is satisfied, we have the unconditional expectation of
sp as follows

Es;l=a,
I( l- ~
i=l
aiE((e,_il-yie,_i)b- f:
j=l
flj

In its special case, when 6 =2 and yi=O, we have the covariance stationarity
condition for E, as

which is the same as that derived by Bollerslev (1986).

When 6 = 2 and yi # 0, we have the covariance stationarity condition for GJR
model as
2-l

~i$l~i[(l+~~)+(1~~i)212 2 ,(y)+ i Bj
j= 1
 Z. Dins ct al., A long mrmory property qfstock market rrturns 105

When 6 = 1 and 7; = 0. we have the condition for existence of Es, and E letI of
Taylor/Schwert model

kit1 zi2r(1)+ i
j=l
Pj

=~ ~ cCi+ ~ Bj<l,
i=l j=l

Since m < 1, so even if cr= 1 pi + x4= 1 /Ii > 1 it can still be true that Es, or
E) E,1exists and is finite, this condition is weaker than the covariance stationarity
condition of the model. It is possible E 1E,1 does not exist and Edis not covariance
stationary even if this condition is satisfied.
When 6 = 1 and ui#O, we have the existence condition of Es~ and E 1~~1 for the
Asymmetric Taylor/Schwert model or the generalized Zakoian model which is
the same as that for the Taylor/Schwert model fi cr= 1 Cli+ x4= 1 bj < 1.
Under the assumption that

i.e. the 6th moment of s, and le,l exist, then if 622 we have that E, is covariance
stationary. If 6 3 1 then Es, and E IE,I exist and are finite. But 6 3 2 is a sufficient
condition for the process E, to be covariance stationary.

References

Black, Fisher, 1976, Studies in stock price volatility changes, Proceedings of the 1976 business
meeting of the business and economics statistics section, American Statistical Association,
1777181.
Bollerslev, T., 1986, Generalized autoregressive conditional heteroskedasticity, Journal of Econo-
metrics 31, 3077327.
Bollerslev, T. and R.F. Engle, 1992, Common persistence in conditional variance, Forthcoming in
Econometrica.
Davidian, M. and R.J. Carroll, 1987, Variance function estimation, Journal of American Statistical
Association 82, No. 400, 1079-1091.
Eatwell, J., M. Milgate and P. Newman (eds.), The new palgrave: Finance (New York, Norton).
Engle, R.F. 1982, Autoregressive Conditional heteroskedasticity with estimates of the variance of
U.K. Inflation, Econometrica, 50, 987-1008.
Engle, R.F. 1990, Discussion: stock volatility and the crash of 87, Review of Financial Studies,
Volume 3, Number 1, 1033106.
Engle, R.F. and T. Bollerslev. 1986, Modelling the persistence of conditional variances, Econometric
Review, 5, l-50, 81-87.
Engle, R.F. David Lilien and Russ Robins, 1987, Estimating time varying risk premia in the term
structure: The ARCH-M Model, Econometrica 55, 391407.
Engle, R.F. and G. Gonzalez Rivera, 1991, Semiparametric ARCH models, Journal of Business and
Economic Statistics 9, 3455360.
106 Z. Ding et al., A long memory property qf stock murket returns

Engle, R.F. and V. Ng, 1992, Measuring and testing the impact of news on volatility, Forthcoming in
Journal of Finance.
Fama, E.F., 1970, Efficient capital markets: A Review of Theory and Empirical Work, Journal of
Finance, 25, 383417.
Fama, E.F., 1976, Foundations of finance: Portfolio decision and security prices, New York: Basic
Books Inc.
French, Ken, William Schwert and Robert Stambaugh, 1986, Expected stock returns and volatility,
Journal of Financial Economics, 19, 3-29.
Glosten, L., R. Jaganathan and D. Runkle, 1989, Relationship between the expected value and the
volatility of the nominal excess return on stocks, unpublished manuscript, J.L. Kellogg Graduate
School, Northwestern University.
Granger, C.W.J., 1991, Forecasting stock market prices: Lessons for forecasters, UCSD Working
Paper
Granger, C.W.J., 1980, Long memory relationships and the aggregation of dynamic models. J. of
Econometrics, 14, 2277238.
Granger, C.W.J. and A.P. Anderson, 1978, An introduction to bilinear time series model, Vanden-
hoek and Ruprecht, Gottingen
Granger, C.W.J. and R. Joyeux, 1981, An introduction to long-memory time series models and
fractional differencing, J. of Time Series Analysis, 1, 15-29.
Granger, C.W.J. and 0. Morgenstern, 1970, Predictability of stock market prices. Heath-Lexinggton
Press.
Granger, C.W.J. and Paul Newbold, 1986, Forecasting economic time series, New York, Academic
Press.
Hamao, Y., R.W. Masulis, V. Ng, 1990, Correlation in price changes and volatility across interna-
tional stock markets, The Review of Financial Studies, V3, 2. 281-307.
Higgins, M. and A. Bera, 1990, A class of nonlinear ARCH models, Working Paper, Department of
Economics, University of Wiscosin at Milwaukee.
Kariya, T. Tsukuda, Y. Maru, J., 1990, Testing the random walk hypothesis for Japanese stock price
in S. Taylors Model, Working paper, University of Chicago.
Nelson, D.B., 1990, Stationarity and persistence in the GARCH(l.1) model, Econometric Theory, 6,
318-334.
Nelson, D.B., 1991. Conditional heteroskedasticity in asset returns: A New Approach. Econo-
metrica, Vol. 59, No. 2, 347-370.
Schwert, W., 1990, Stock volatility and the crash of 87, Review of Financial Studies, Vol. 3, No. 1,
777102.
Taylor, S., 1986, Modelling financial time series, New York, John Wiley & Sons.
Zakoian, J., 1991, Threshold heteroskedasticity model, unpublished manuscript, INSEE.