A Comparison of Volatility Predictions in the

Hong Kong Stock Market

Law, Ka Chung

School of Economics
University of Hong Kong

June 1999

Master of Economics
Dissertation (M6003)
 ABSTRACT

With the introduction of new financial instruments in recent years especially a variety

of derivative securities, financial markets have become more complex and, to certain

degree, more volatile. Volatility prediction has thus become more important for both

practitioners and academics. Using only historical data, this paper examines a number

of existing volatility predicting models. Among them, the Random Walk model, the

GARCH model, the EGARCH model and the Stochastic Volatility model are

examined with certain modifications. In addition, Hang Seng Index Option prices are

used as an instrument for analysis.

Many thanks to my supervisor, Dr. G. G. Yu, who kindly guided my

writing of this paper, especially for his valuable advice.

Many thanks to the assistant computer officer of the School of Economics

and Finance, Miss Lee Shuk Fong, who helped a lot with great effort in

the empirical part of this paper.

Without their support this paper cannot be finished successfully.

Introduction 1

Traditional Models 3

Volatility Predictors 7

Actual Volatilities 12

The Hang Seng Index Option 15

Methodology 17

Results 22

Interpretation 37

Conclusion 53

Appendix 56

Note 57

Bibliography 60
 HSI Trend

18000

16000

14000

12000

10000
HSI

8000

6000

4000

2000

0
01/03/84 15/10/84 29/05/85 10/01/86 26/08/86 09/04/87 23/11/87 06/07/88 17/02/89 03/10/89 17/05/90 31/12/90 14/08/91 27/03/92 10/11/92 24/06/93 07/02/94 21/09/94 05/05/95 19/12/95 01/08/96 17/03/97 29/10/97 12/06/98 27/01/99
Time
A Comparison of Volatility Predictions in the Hong Kong Stock Market

INTRODUCTION

Volatility prediction is of the most important topics in financial

studies. Not only volatility estimation is crucial for active traders to make

optimal trading decision, but it also helps passive investors to set their

portfolios consistent with their risk exposure. With the recent introduction of

a wide range of financial derivatives where short selling are made possible

through synthetic options and transaction costs could be reduced, it is

interesting to examine different volatility prediction models using both the

trading data of index and index options. Given Hong Kong’s fame of a very

volatile market, this paper makes the first step of tackling the volatility

estimation for the Hong Kong stock markets.

A glance of the Hang Seng Index (HSI) chart in the last 15 years will

help get the rough idea. The HSI trend shows the increasing fluctuation since

about the mid-1992. Notice before that, since about the mid-1991, the HSI has

already started rising. It might have been the case that the rise of the index

had led to the increase in the volatility. This is the trivial result of the well-

known Capital Asset Pricing Model (CAPM), which relates the return to risk

in a positive way.1

Nevertheless, simply by knowing the positive co-movement between

the two is not enough for conducting volatility prediction. Nor simply having

1
A Comparison of Volatility Predictions in the Hong Kong Stock Market

a linear CAPM model will do. One of the main reasons is the recent

introduction and later proliferation of the financial derivatives. As mentioned

earlier, the multiplier nature of such instruments essentially may have played a

role in the rapid rise and fall of the HSI, and hence the corresponding Hong

Kong stock market volatility. This kind of multiplier effect is so complicated

that linear CAPM is unlikely to be able to explain.

This paper will focus on some other non-linear return-volatility

models. Although they do not explicitly model the relationship between the

market volatility and return, the aim is just in line with that of the CAPM.

That is, given a set of return data, how would the corresponding volatility look

like? Since the objective of this paper is essentially empirical, the theoretical

models are not discussed in detail. Rather, they are examined using the index

options, and the Hong Kong stock market data in the past decade.

The reason for doing because index option is one of the most

representative financial derivatives, which have been contributing much

change to the stock market volatility structure since its introduction. Where

Hong Kong is chosen is simply for convenience and its relative maturity of the

financial market among other regional countries. In the following sections, we

will first summarize the popular volatility models, then we describe the

models used in this paper, and then we present the methodology, empirical

results and our interpretations for the findings.

2
A Comparison of Volatility Predictions in the Hong Kong Stock Market

TRADITIONAL MODELS2

Before going into the specifics of this paper, here the traditional

approaches to the volatility prediction will be briefly summarized.

In general the objective of volatility prediction is to obtain the

expected future volatility, for two reasons. First, it is the important input for all

dynamic trading and option pricing models, which are the generalized results

of the Black-Scholes option pricing model (1973). Kat (1993a, 1993b) shows

that more accurate volatility predictions can greatly improve the replication of

the efficiency of delta hedging strategies by using the Black-Scholes hedge

ratios. Also, the expected volatility is also important for static hedging and

portfolio setting, etc..

There are different kinds of volatility predictors for which expected

future volatility can be obtained. Some are based directly on the historical

return data; some are based on the observed option prices with the

corresponding option pricing models to obtain the volatility estimates; while

some are based on multi-factor models. In addition, some combinations of

them are possible and may perform even better. In this paper the predictors

used are based on the historical returns only. Within this class of predictors,

there are again several kinds of approaches available.

3
A Comparison of Volatility Predictions in the Hong Kong Stock Market

One of them is to model the time series behaviour of volatility using a

two-stage method explicitly. Here a series of sample volatility estimates over

time are obtained first, and then are fitted into the standard AutoRegressive

Integrated Moving-Average (ARIMA) models.3 Notice that this approach

assumes the constancy of volatility within the period where data are used.

Moreover, this approach reduces the amount of information available. Hence

the variance of the volatility estimates in the ARIMA model increases. Also,

Chou (1988) shows that the estimates are extremely sensitive to the sampling

frequency.

Another approach is still to predict volatility from historical return

data, but are treated only as part of a model of the return behaviour. There are

typically four of them. The first is the Random Walk model. It assumes the

constancy of the volatility (homoscedasticity), even though this is empirically

implausible. From the graph discussed earlier we can see that the volatility is

unlikely to be constant over time. Unless there is great estimation error on the

return data, the constant volatility hypothesis is unlikely to be accepted.4

To overcome this, models that allow changing volatility

(heteroscedasticity) are then needed. The AutoRegressive Conditional

Heteroscedasticity (ARCH) class models are popular in handling this, where

volatility is a deterministic function of historical returns. The ARCH model

4
A Comparison of Volatility Predictions in the Hong Kong Stock Market

was first invented by Engle (1982), where the conditional disturbance is

expressed in terms of the lagged disturbances in polynomial form:

ARCH (q): Et(εt+12) = a0 + a1 εt-12 + … + aq εt+1-q2

Later, Bollerslev (1986) added the conditional variance terms to develop the

Generalized ARCH – the GARCH model:

Then Nelson (1991) modified it into the log form as the Exponential GARCH

– the EGARCH model:

where εt is white noise disturbance, ht is the conditional standard deviation,

ai’s and bi’s are constants. These three of the ARCH class models did quite

well in the description of stock price dynamics.

Apart from this class, there are models where the conditional variance

depends not only on the past variances but also on the random noise. The

continuous time stochastic stock return volatility model developed by Hull and

White (1987) which prices stock options is one example.

In this paper, the Random Walk, GARCH, EGARCH and Stochastic

Volatility models will be used in volatility comparison. The inadequacy of

using ARIMA model is mentioned above and hence discarded. While the

5
A Comparison of Volatility Predictions in the Hong Kong Stock Market

giving up of the ARCH model is because its similarity compared to the

GARCH and EGARCH ones, with the latter two in general perform much

better (as much more relevant past information is used). Since the EGARCH

and Stochastic Volatility models are not commonly used as the GARCH one

in the volatility prediction context, in this paper the EGARCH and the

Stochastic Volatility models will be merged into one. Why is and how this can

be done will be discussed in the next section.

6
A Comparison of Volatility Predictions in the Hong Kong Stock Market

VOLATILITY PREDICTORS

In this section the details of the models will be explored in turn.

The return of all the models is commonly given by:

Rt = Et(Rt) + εt = µ + ht vt …………………………………………...……... (1)

where Rt = ln(Pt / Pt-1) = (Pt – Pt-1) / Pt = Return, Pt = Index;5

ht = Et-1(εt2) = Conditional standard error (or simply volatility),

σt = E(εt2) = Unconditional standard error,

εt = ht vt = Disturbance, vt ~iid N(0,1) = White noise process.

In our context the expected return is assumed to be constant (µ). In contrast to

return, the functional form of the volatility differs in different models.

1. Random Walk Model

The conditional variance of the Random Walk (RW) model is:

ht = ⎯σ …..………………………...…………...………………………….... (2)

Clearly, the conditional standard error (volatility) is simply the unconditional

standard error, which is constant over time by construction. Given this

probabilistic model, the estimate should be obtained from as many

observations as possible to reduce the sampling error. Since observations are

assumed to be independent and identically distributed, the average daily

volatility estimate (the RW predictor, denoted by h2Av) over T days of

prediction since date t, by using t of historical daily observations is simply:

7
A Comparison of Volatility Predictions in the Hong Kong Stock Market

h2Av(t,T) = σt2 T ………………………………………………………….... (3)

1 t 1 t
σ t2 = ∑
t − 1 τ =1
( Rτ − ∑ Rτ ) 2 …………………………………………….… (4)
t τ =1

where t is the total number of observations, τ = 1, 2, …, t.

2. GARCH Model

Here only the simplest form, the GARCH (1,1), will be considered.

The conditional variance is given by:

ht2 = α0 + α1 εt-12 + β1 ht-12 …………………………………………………. (5)

where γ = α1 + β1 < 1 and α0, α1, β1 > 0. Also, σt =⎯σ.

That is, the conditional variance (ht2) is time-dependent even though the

unconditional one (σt2) is not. Notice that returns (Rt) tend to be large in

magnitude successively, due to the autocorrelated relationship of the

conditional variance. That is, large ht-12 implies large ht2 [by (5)], even though

they may be of different signs. Hence by (1), large Rt-12 (or |Rt-1|) implies large

Rt2 (or |Rt|).

Note also that although the conditional standard deviation (ht) is

normal, the unconditional one (σt) is leptokurtic (i.e., flat-tailed). By repeated

substitution of (5), as shown the Appendix given by Heynen, Kemna and

Vorst (1994), the GARCH average volatility predictor (daily based) over T

days after date t is given by:

8
A Comparison of Volatility Predictions in the Hong Kong Stock Market

2 2 1− γ T
h (t , T ) = σ + (h
2 2
t +1 −σ ) …………….………………………….. (6)
T (1 − γ )
Av

2 α0
σ = ………………………………………………………………… (7)
1−γ

Here σ is the mean of σt, and γ can be interpreted as a measure of the speed

by which shocks to volatility decay. For γ approaches 1 from below (γ → 1-),

meaning the effect of past volatility increases, the results are two-fold. As T

increases, the predictor is mean reverting; for T decreases, it converges to the

next period volatility:

1−γ T
For large T: lim = lim γ T −1 = 0 ⇒ h2Av(t,T) = ⎯σ2 → +∞ [by (7)]
γ →1 T (1 − γ ) γ →1

1−γ T
For small T: lim = lim γ T −1 = 1 ⇒ h2Av(t,T) = ht+12
γ →1 T (1 − γ ) γ →1

In a word, two factors are responsible for the mean reversion: significant effect

of past information and long predicting period.

3. Mixture: EGARCH and Stochastic Volatility Models

In the above models the average volatility predictor [h2Av(t,T)] is

derived within the same model. Here it is derived from two different models.

Nelson (1990) shows that the EGARCH (1,1) is a good approximation to the

discrete time AR (1) process; where in continuous time limit, the EGARCH

also converges to the same AR (1) process. On the other hand, Dassios (1992)

shows that both EGARCH (1,1) and AR (1) converge to the popularly used

9
A Comparison of Volatility Predictions in the Hong Kong Stock Market

continuous time Stochastic Volatility (SV) model given by Hull and White

(1987). Despite the latter converges faster than the former, the approximation

is valid for large t. The logic is shown as follows:

EGARCH (1,1) → Discrete AR (1) → Stochastic Volatility

Compare the EGARCH with the SV model, the SV model has

features significantly different from the GARCH one. Also, the average

volatility predictor given by the SV model is much simpler (while that of the

EGARCH is very complicated). Therefore, the average volatility predictor of

the SV model will be used. Since the log-likelihood function and hence the

volatility formula of the SV model cannot be expressed explicitly, an

approximation to the continuous version of the SV model, the EGARCH

model, is used in generating relevant parameters for volatility prediction.

However, the data obtained are in discrete form (i.e., daily basis), the

discrete AR(1) process has to be used to approximate the EGARCH model in

generating the parameters. All these procedures can be justified by the

arguments in the previous paragraph. The conditional variance of the

EGARCH (1,1) model is given by:

2
ln ht2 = α 0 + α 1 ln ht2−1 + β 1vt −1 + φ ( vt −1 − ) …………………………...…. (8)
π

where α0, β1, β2 and φ are all independent parameters which may take on any

values. This is because the conditional variances (ht2) are in log-form, which

10
A Comparison of Volatility Predictions in the Hong Kong Stock Market

may take on any value; while those in the GARCH model must be non-

negative, and hence the restrictions on its parameters [see (5)]. The last term in

right hand side of (8) allows the capturing of asymmetric changes in the index

by the conditional variance.

The AR (1) process used to approximate the EGARCH (1,1) is:

ln(ht2) = α0 + α1 ln(ht-12) + ut ………………………………………………. (9)

where ut ~iid N(0,σu2) and E(ut vt) = 0.

For stochastic volatility model, log-likelihood function cannot be specified

explicitly. To obtain its volatility predictor, three approaches are possible.

Simulation methods can be used but the computations are very tedious.6

Another way is to use the generalized methods of moments to obtain the

maximum likelihood parameter.7 It can also be obtained by the Kalman filter

technique, where Ruiz (1993) shows that the estimates are the most efficient

among 3 approaches. Heynen and Kat (1994) use the Kalman filter technique

to obtain the average volatility predictor (of the SV model):

2
σ T
α 0α 1k σ u2α 12 k
h (t , T ) =
2
Av
T
∑h
k =1
t
2α1k
exp(− −
1 − α 1 2(1 − α 12 )
) ………………….………. (10)

2 α σ2
where σ = exp( 0 + u
) . It can be shown that:
1 − α1 2(1 − α12 )

For large T: h2Av(t,T) → ⎯σ2

For small T: h2Av(t,T) → ht2

This is obviously consistent with that given by the GARCH model.

11
A Comparison of Volatility Predictions in the Hong Kong Stock Market

ACTUAL VOLATILITIES

To evaluate the predictability of the above models, actual volatility is

used as a standard for comparison. In this paper two kinds of actual volatility

will be used for evaluation.

1. Historical Volatility

Historical Volatility (HV) is simply the standard deviation of the past

returns. The average volatility prediction by the above models over the period

from date t+1 to date t+T [t+1,t+T], which utilizes the information up to date t

(τ∈[1,t]), should be compared to the HV over the same period [t+1,t+T]. It is

also the most popular measure used in practice:

1 t +T 1 t +T
2
s HV (t , T ) = ∑
T − 1 j =t +1
(R j −
T
∑R )
j =t +1
j
2
…………..……………………….. (11)

Notice that (11) has already incorporated the effect of “average”, which is

comparable to the hAv2(t,T) given earlier.

2. Implied Volatility

Since the index options traded in HK are of European type, the Implied

Volatility (IV) is given by the Black-Scholes option pricing formula:

C = S e-γT N(d1) – X e-rT N(d2),

P = X e-rT N(–d2) – S e-γT N(–d1);

12
A Comparison of Volatility Predictions in the Hong Kong Stock Market

ln( S / X ) + (r − γ + σ 2 / 2)T
where d1 = , d 2 = d1 − σ T ………………… (12)
σ T

N(.) = Cumulative probability function for a standardized normal variable;

C = Call price, P = Put price, S = Index price, X = Strike price, r = Risk-free

interest rate, T = Maturity, γ = Dividend yield, σ = Volatility of stock price.

Traditionally the above formulae are used for individual stocks. In the

index option context, variables have to be modified to the index context.

Given the time series data on all other variables σ can be obtained easily, and

hence the corresponding “average” IV [sIV2(t,T)] can be obtained. In fact, the

tailor-made time series of data are available from the Hong Kong Futures

Exchange (HKFE).8 Therefore, we can ignore the detail mechanism of how the

IV series are obtained.

Before leaving this section a note is worthy mentioning. Both kinds of

volatility have to be obtained with the index as the ultimate input via its return

However, the time intervals between successive (close-to-close) observations

are not always restricted to one day, which is due to the presence of weekends

and holidays. Fortunately, stock return (and hence index return) volatility is

rather low in the non-trading days in general. French and Roll (1986) find that

in US the close-to-close return volatilities over weekends are only about

10.7% higher than those in the trading days.

13
A Comparison of Volatility Predictions in the Hong Kong Stock Market

For HK, the HSI close-to-close return volatilities over both weekends

and holidays are 14.5% lower than those in the trading days.9 This presents an

acceptable level of difference which follows that the weekends and holidays

effect is negligible. Therefore, in this paper all the time intervals between

successive observations will be treated as identical – 24 hours.

14
A Comparison of Volatility Predictions in the Hong Kong Stock Market

THE HANG SENG INDEX OPTION

In contrast to the futures markets, the options market in HK has not

become popular until very recently. A few individual stock options emerged a

little earlier but trading were very thin. Because of that, index option only acts

as an informal “representative” to the option market. There are 4 index options

available in HK: a) Hang Seng Index Options (HSIO); b) Hang Seng 100

Options; c) Red-Chip Options; and d) HKFE Taiwan Index Options. Apart

from the first one, the other three were introduced only very recently (within

last 2 years); and their volumes have been relatively small compared to the

HSIO.10 This is the main reason why the HSIO is chosen for analysis in this

paper.

The HSIO contracts were first introduced on 5th March 1993.

Obviously, the underlying asset of the HSIO is the HSI, compiled, computed

and disseminated by the HSI Services Ltd. The contract multiplier is HK$50

per index point, with the spot, next 2 calendar months, next 3 calendar quarter

months and next 2 months of June and December as contract months. The

minimum fluctuation is 1 index point; no position limits are set, and 500

contracts in a series are considered to be large open positions; the cabinet bids

are HK$10 per contracts.

15
A Comparison of Volatility Predictions in the Hong Kong Stock Market

Up to now open outcry is practiced. HSIOs are European style which

can only be exercised on the expiry day, which is the business day

immediately preceding the last business day of the contract month. Settlement

is done through cash at the final value, and the final settlement day is the

business day immediately following the expiry day.

Since it is not the purpose of this paper to focus on the mechanism of

the HSIO. We will now turn to the methodology section.

16
A Comparison of Volatility Predictions in the Hong Kong Stock Market

METHODOLOGY

Since there has been research in comparing the index return volatilities

across markets in different countries,11 in this paper we will compare within

the local market only. In the followings the mechanism of the practical work

will be discussed.

Two kinds of data are needed in the first place: the HSI prices and their

actual volatilities. The data are divided into 2 consecutive sub-periods. The

first (information period) is for prediction purpose, where the volatility

predictors (i.e., the expected future volatility estimates) are obtained from the

HSI prices via the 3 models discussed in the “Volatility Predictors” section

respectively. The second (evaluation period) is for evaluation, where the

estimates obatined are compared to the actual volatilities derived from the

“Actual Volatilities” section to see if they are good predictors. Also, they will

be compared to each other to see which is the best are.

Index returns (Rt) and their corresponding unconditional variances (σt)

are easy to obtain by simply using the HSI with the formulae given in the

previous section. This can be done through many common spreadsheet

programs such as Excel. While the conditional variances (ht) have to be

obtained through some specific software (in this paper SHAZAM and Microfit

are used). After these data have been obtained, the regression parameter

17
A Comparison of Volatility Predictions in the Hong Kong Stock Market

estimates (α’s and β’s) and hence the volatility predictors [h2Av(t,T)] can be

mechanically calculated by any statistical programs.

The HSI data are available from 1984 up to now, but those of the

HSIO are available only from 5th March 1993 to 28th February 1999 (almost 6

years). Since both kinds of data are essential for analysis, only the period

where data of both kinds are available are used in our context (i.e., 5/3/1993 –

28/2/1999). The HSI data are used to generate the return, unconditional and

conditional variances and HV, with everything expressed in daily terms. The

HSIO data are given with the IV series, which are calculated from the average

settlement prices of the 6 at-the-money series (3 calls & 3 puts) by the HKFE.

In the first stage, different lengths of information periods have been

tried. The “daily” return from 5th to 8th March 1993 (since 6th and 7th are

weekends) is set as the initial point of the data series. The final point is then

the “daily” return from the second last to the last trading day of each year from

1994 to 1998 are tried (5 trials in total). Given the relationship between the

daily return and the HSI prices is Rt = (Pt – Pt-1) / Pt , the HSI at 5/3/93 is P0, at

8/3/93 is P1, …, Pt is then the last trading day in December of 1994, 1995, …,

1998 respectively. Similar notion applies to the daily return, such as R1

denotes the daily return from 5/3/93 to 8/3/93.

18
A Comparison of Volatility Predictions in the Hong Kong Stock Market

The estimates obtained from 1993 – 1994 and 1993 – 1995 periods

are awkward. That is, the estimates show irregularity for too short information

periods, such as γ > 1 for the GARCH model. So, these two trials must be

discarded.

In the second stage, the information periods are refined with more

“final days”. Apart from the last trading day of each year from 1996 to 1998

(i.e., 3 different information periods), the last trading days of each quarter are

tried, i.e., the end of March, June and September. That means there are 12 (=

3×4) trials totally. Notice that all the “final days” are simply date t in terms of

the previous notation.

In the third stage the volatility predictions are evaluated by comparing

the HV’s and the IV’s [si(t,T), i = HV or IV] with the average volatility

predictors [hAv(t,T)] of the 3 models respectively. That is, there are 6

combinations in total. The period from t+1 to t+T just after the information

period (t) is the evaluation period. By saying “comparing volatilities in the

evaluation period of length T”, we mean both average “predicted” and “actual”

volatilities series from t+1 to t+2, t+1 to t+3, …, t+1 to t+T are obtained (T of

them in total) and compared. For inter-model comparison (i.e., to find which is

the best), both of the compared series are average predicted volatilities.

19
A Comparison of Volatility Predictions in the Hong Kong Stock Market

For each model, T is set to be 10, 25, 50 and 100 respectively.12 The

reasons for setting these numbers are as follows. First, the interval between

any 2 “final days” of successive information periods is less than 70 days.13

Hence setting T at the maximum of 100 (long horizon forecast) leaves no

trading days unpredicted. We will see in later sections that setting T above 70

has its significance, as which model performs the best depends heavily on T.

Second, setting T lower than 10 may cause the evaluation inaccurate due to the

too small sample size. Therefore, T = 10 (about 2 weeks) is the shortest

horizon forecast possible. In between are the medium forecasts where T = 25

and T = 50 represent 1 and 2 month(s) forecasts respectively.

Concern only the data series with T = 100, regress the HV and IV

series on the average predicted volatility series generated by the 3 models

respectively (i.e., 6 regressions for each evaluation period from t+1 to t+T).

The corresponding coefficient of determination (R2) is the measure of the

forecast ability. Then F test can be applied (adjusted with appropriate degrees

of freedom) to test the predictability hypothesis. Nonetheless, as shown in the

next section, the R2 varies quite a lot from sample to sample such that the

“0.7–rule” can be used, i.e., good prediction for R2 > 0.7. Notice that only T =

100 is concerned since regression cannot be applied to small sample (T < 30).

Although regression is a popular method for evaluation, it cannot

evaluate the predictability wherever T is small. Thus, mean squared deviation

20
A Comparison of Volatility Predictions in the Hong Kong Stock Market

(MSD) between the average predicted and the actual volatilities will also be

used as the standard for comparison among the volatilities predicted by the 3

models. Its expression is given by:

T
[h Avj (t , Γ) − si (t , Γ)]2
MSD (h, s ) = ∑
Γ =t +1 T
……………………………………. (13)

where i = HV or IV, j = RW, GARCH or Mixed.

Notice that in this approach (MSD) statistical inference (such as F

test) cannot be made, but comparisons can be still be done at some lower level.

For instance, the lower the MSD, the better the prediction. As our another

objective is also to find the best model among the three, another form of the

MSD will also be used in our context, which is given by:

T
[h Avj (t , Γ) − h Av
k
(t , Γ)]2
MSD (h, s ) = ∑ ………………………………..…. (14)
Γ =t +1 T

where j, k = RW, GARCH or Mixed, and j ≠ k.

21
A Comparison of Volatility Predictions in the Hong Kong Stock Market

RESULTS

The first results presented are the parameter estimates. The date given

(mm/yy) refers to the final trading date “t”. Some of the boxes are labeled with

“N/A” (not applicable) which is due to the unavailability of the relevant data.

Denote the average predicted volatility as APV.

1. Parameter Estimates:

Table 1: RW

2 2 2
Date σ Date σ Date σ

3/96 0.000247648 3/97 0.000207727 3/98 0.000343843

6/96 0.000234243 6/97 0.000206063 6/98 0.000352804

9/96 0.000222003 9/97 0.000216574 9/98 0.000375772

12/96 0.000213658 12/97 0.000302495 12/98 0.000385518

22
A Comparison of Volatility Predictions in the Hong Kong Stock Market

Table 2(a): GARCH

Date α0 α1 β1

3/96 3.22E-06 0.060637 0.92664

6/96 3.19E-06 0.069071 0.91669

9/96 2.63E-06 0.065263 0.92209

12/96 8.34E-07 0.50843 0.46265

3/97 2.57E-06 0.059706 0.927

6/97 2.78E-06 0.060524 0.92575

9/97 3.44E-06 0.072373 0.91245

12/97 2.63E-05 0.4427 0.36964

3/98 5.14E-06 0.11174 0.87235

6/98 4.68E-06 0.11071 0.87774

9/98 4.29E-06 0.10923 0.88175

12/98 4.27E-06 0.10757 0.88304

23
A Comparison of Volatility Predictions in the Hong Kong Stock Market

Table 2(b): GARCH (continued)

Date γ σ
2
ht+12

3/96 0.987277 0.000252928 0.000299408

6/96 0.985761 0.000224033 8.23246E-05

9/96 0.987353 0.000208113 7.67544E-05

12/96 0.97108 2.88441E-05 0.000148928

3/97 0.986706 0.00019317 0.000117654

6/97 0.986274 0.00020239 0.000252805

9/97 0.984823 0.000226791 0.000400582

12/97 0.81234 0.000140398 0.000414661

3/98 0.98409 0.000323193 0.000239618

6/98 0.98845 0.000405195 0.000921848

9/98 0.99098 0.000476053 0.000800105

12/98 0.99061 0.000454558 0.000246877

24
A Comparison of Volatility Predictions in the Hong Kong Stock Market

Table 3(a): Mixed

Date α0 α1 σ u2

3/96 -0.135275027 0.98400282 0.010477772

6/96 -0.132573018 0.984600965 0.011467986

9/96 -0.109111103 0.987469396 0.01044875

12/96 -0.114748576 0.986768469 0.009801512

3/97 -0.114079421 0.986884007 0.009311962

6/97 -0.125895921 0.985446496 0.009688548

9/97 -0.175044659 0.979662189 0.014403527

12/97 -0.185393512 0.978192316 0.023058955

3/98 -0.15845644 0.981269335 0.02305279

6/98 -0.223639756 0.973384298 0.032123468

9/98 -0.132590008 0.984048923 0.022550606

12/98 -0.132018715 0.98413041 0.021795104

25
A Comparison of Volatility Predictions in the Hong Kong Stock Market

Table 3(b): Mixed (continued)

Date σ
2
ht2

3/96 0.000250735 0.000263106

6/96 0.000220073 8.88735E-05

9/96 0.000203918 6.92686E-05

12/96 0.000206351 0.000148054

3/97 0.000199625 0.000136464

6/97 0.000206976 0.00024377

9/97 0.000218661 0.000333243

12/97 0.000265463 0.000515558

3/98 0.000288985 0.000260834

6/98 0.000304529 0.000678927

9/98 0.000350547 0.000792908

12/98 0.000344684 0.000312209

26
A Comparison of Volatility Predictions in the Hong Kong Stock Market

2. R2:14

Table 4: R2 between HV and RW, GARCH, Mixed APV

Date Random Walk GARCH Mixed

3/96 0.190564 0.214334 0.22228

6/96 0.065293 0.060872 0.055579

9/96 0.758935 0.758854 0.747625

12/96 0.20645 0.138545 0.297323

3/97 0.184833 0.254666 0.24792

6/97 0.758499 0.800865 0.678884

9/97 0.003946 0.014454 0.018083

12/97 0.320102 0.363986 0.002759

3/98 0.867496 0.739821 0.855212

6/98 0.56574 0.591587 0.592616

9/98 0.694584 0.754379 0.737711

12/98 0.155823 0.179944 0.148666

0.9
0.8
0.7
0.6
0.5 RW
0.4 GARCH
0.3 Mixed
0.2
0.1
0
3/96
6/96
9/96
12/96
3/97
6/97
9/97
12/97
3/98
6/98
9/98
12/98

27
A Comparison of Volatility Predictions in the Hong Kong Stock Market

Table 5: R2 between IV and RW, GARCH, Mixed APV

Date Random Walk GARCH Mixed

3/96 0.510276 0.157399 0.552282

6/96 0.010373 0.009346 0.005075

9/96 0.634125 0.638607 0.641066

12/96 0.473616 0.392313 0.53919

3/97 0.323294 0.315774 0.310103

6/97 0.539345 0.570243 0.497037

9/97 0.00158 0.002697 0.008971

12/97 0.493631 0.452652 0.27979

3/98 0.476595 0.537542 0.507397

6/98 0.077252 0.026588 0.048274

9/98 0.173246 0.203437 0.196233

12/98 0.125546 0.128069 0.124805

0.7

0.6

0.5

0.4 RW
0.3 GARCH
Mixed
0.2

0.1

0
3/96
6/96
9/96

3/97
6/97
9/97

3/98
6/98
9/98
12/96

12/97

12/98

28
A Comparison of Volatility Predictions in the Hong Kong Stock Market

3. MSD – HV:

Table 6: MSD between HV and RW, GARCH, Mixed APV

Table 6(a): HV for T = 10 (Unit: 10-8)

Date Random Walk GARCH Mixed

3/96 20.83686 0.425008 0.265688

6/96 18.99482 0.030611 0.028159

9/96 17.30721 0.006916 0.004842

12/96 14.19155 0.006876 0.259791

3/97 10.99344 0.366756 0.314707

6/97 13.79727 0.22891 0.198222

9/97 11.56655 0.372102 0.128236

12/97 6.440784 8.287989 7.923379

3/98 39.97928 0.129932 0.184178

6/98 29.25276 1.24832 0.27271

9/98 14.13871 3.067227 3.89136

12/98 34.80402 1.194303 0.60045

29
A Comparison of Volatility Predictions in the Hong Kong Stock Market

Table 6(b): HV for T = 25 (Unit: 10-8)

Date Random Walk GARCH Mixed

3/96 130.2592 0.446194 0.316765

6/96 115.9852 0.013121 0.012701

9/96 105.1408 0.00649 0.003329

12/96 92.70272 0.014926 0.707685

3/97 85.16423 0.179824 0.141889

6/97 87.61585 0.186889 0.169922

9/97 18.28475 41.78188 47.48098

12/97 66.54294 33.74375 15.86715

3/98 246.9795 0.097933 0.149547

6/98 234.6215 1.527634 0.241887

9/98 20.68925 3.671214 4.333184

12/98 48.16279 1.286862 0.576432

30
A Comparison of Volatility Predictions in the Hong Kong Stock Market

Table 6(c): HV for T = 50 (Unit: 10-8)

Date Random Walk GARCH Mixed

3/96 517.1973 0.461405 0.350275

6/96 461.9645 0.01506 0.015836

9/96 415.2903 0.009201 0.003943

12/96 378.7266 0.008307 2.975283

3/97 354.1654 0.091981 0.072262

6/97 340.6466 0.182182 0.17787

9/97 181.2503 50.0554 52.85845

12/97 539.169 29.96853 9.799129

3/98 977.4698 0.060969 0.096483

6/98 951.8851 1.01207 0.432479

9/98 30.37574 4.081735 4.621164

12/98 N/A N/A N/A

31
A Comparison of Volatility Predictions in the Hong Kong Stock Market

Table 6(d): HV for T = 100 (Unit: 10-8)

Date Random Walk GARCH Mixed

3/96 2078.392 0.456797 0.368442

6/96 1858.686 0.031322 0.03171

9/96 1662.776 0.006568 0.002173

12/96 1533.985 0.013406 17.72175

3/97 1439.656 0.050206 0.03745

6/97 1303.772 2.674311 2.758318

9/97 1189.497 38.64866 40.75698

12/97 2783.568 18.92021 50.9466

3/98 3888.932 0.30282 0.293122

6/98 3959.678 0.594021 1.148214

9/98 43.24509 4.343589 4.888267

12/98 N/A N/A N/A

32
A Comparison of Volatility Predictions in the Hong Kong Stock Market

4. MSD – IV:

Table 7: MSD between IV and RW, GARCH, Mixed APV

Table 7(a): IV for T = 10 (Unit: 10-8)

Date Random Walk GARCH Mixed

3/96 18.61261 0.118198 0.059053

6/96 18.4201 0.006221 0.004058

9/96 16.37758 0.00752 0.012753

12/96 12.68576 0.079178 0.073961

3/97 13.07291 0.034431 0.018045

6/97 11.99932 0.02079 0.016774

9/97 8.887015 0.086357 0.186344

12/97 2.871186 14.21621 9.523105

3/98 33.04145 0.249807 0.161669

6/98 24.495 0.154126 0.417345

9/98 22.82455 0.876953 0.974896

12/98 32.9454 1.513289 0.737019

33
A Comparison of Volatility Predictions in the Hong Kong Stock Market

Table 7(b): IV for T = 25 (Unit: 10-8)

Date Random Walk GARCH Mixed

3/96 125.6223 0.167239 0.103374

6/96 113.9712 0.010306 0.008309

9/96 102.4328 0.007279 0.01301

12/96 88.57147 0.110005 0.376029

3/97 87.44508 0.020238 0.010736

6/97 83.67549 0.024854 0.021457

9/97 30.79902 17.83114 20.67851

12/97 104.4516 17.66594 6.121125

3/98 219.4688 0.462839 0.132778

6/98 206.2143 0.227073 0.330749

9/98 33.16275 0.878805 0.956348

12/98 45.39955 1.634806 0.679785

34
A Comparison of Volatility Predictions in the Hong Kong Stock Market

Table 7(c): IV for T = 50 (Unit: 10-8)

Date Random Walk GARCH Mixed

3/96 508.8445 0.204312 0.139903

6/96 457.5301 0.007275 0.006349

9/96 409.2535 0.008377 0.014926

12/96 371.0023 0.089505 2.334453

3/97 351.5775 0.025111 0.014405

6/97 325.1433 0.291145 0.328054

9/97 280.4888 14.99184 16.25689

12/97 629.5258 13.06599 8.316239

3/98 887.346 1.968273 0.112948

6/98 873.4093 1.579547 0.251271

9/98 46.7296 0.859134 0.892152

12/98 N/A N/A N/A

35
A Comparison of Volatility Predictions in the Hong Kong Stock Market

Table 7(d): IV for T = 100 (Unit: 10-8)

Date Random Walk GARCH Mixed

3/96 2066.231 0.25165 0.193881

6/96 1848.924 0.008056 0.007742

9/96 1642.602 0.045142 0.061988

12/96 1522.719 0.087561 16.61516

3/97 1426.55 0.080505 0.062519

6/97 1231.708 8.826071 8.857107

9/97 1332.427 13.19709 14.3019

12/97 2913.717 7.660893 61.51213

3/98 3727.211 3.552242 0.104682

6/98 3954.58 1.008296 0.141892

9/98 65.0443 0.799033 0.866339

12/98 N/A N/A N/A

36
A Comparison of Volatility Predictions in the Hong Kong Stock Market

INTERPRETATION

1. Parameter estimates

Table 1 shows the trend of the unconditional variance (⎯σ2) over the

investigated period, which is also the conditional variance (ht2) under the RW

model. Recall that for all of the 12 figures shown, the starting date of each data

set (which is used to obtain the variance) is the same – the daily return from

5/3/1993 to 8/3/1993. While the “date” shown in the table is the ending date of

a particular data set, for instance, “3/96” means that the data set ends at the last

trading day of March 1996.

It is clear from the table that the variance is decreasing from 3/96 to

6/97, but is then increasing from 6/97 to 12/98. The reason for the decrease in

the first stage is because of the data accumulation. From “3/96” to “6/97”, the

length of the information period increases from 3 years to 4.25 years.

Statistically, as more data points have been gathered, the variation of the data

set becomes smaller. This is given by the well-known Central Limit Theorem.

However, the Asian Financial Crisis (AFC) happened in the mid-1997

(just after the transfer of sovereignty in July) made the HSI market very

volatile. Thus, the volatility started increase from the period just after the

starting of the AFC (9/97), where its effect on the volatility outweighs that by

the data accumulation. Note also that the increase is the most dramatic from

37
A Comparison of Volatility Predictions in the Hong Kong Stock Market

9/97 to 12/97 and is the second most from 12/97 to 3/98, which are consistent

with the above explanation. The change (both increase and decrease) in other

times are more or less stable.

Table 2(a) presents the parameters of the GARCH model. Since none

of the constant terms (α0) are negative and all of the α1+β1 pairs are less than

1, the well-defined process is guaranteed. Although the constant term does not

have much meaning, we can still infer something from the remaining two

parameters. α1 measures the contribution of the noise in the past period (εt-12)

to the conditional variance in the current period (ht2). The parameter is

generally stable (around 0.06 – 0.07) from 3/96 to 9/97 (except 12/96 15), and

is also very stable from 3/98 to 12/98 (around 0.11). That is, over these 2

stages the contribution of past noise to conditional variance has increased by

roughly 70%. While the sudden increase in the parameter value (to 0.44) in

12/97 can be explained by the AFC, which gave much external shock

(volatility) to the market. This “residual effect” of the AFC can be seen from

the 70% increase in the noise part. Treating α1+β1 as “almost” constant sum

(at least close to the upper bound “1”), the intuition of β1 is just symmetric but

opposite to that of α1.

Table 2(b) gives justification to the above statement (just the last

sentence). It can be seen that the sum (γ) is almost constant (around 0.98 –

38
A Comparison of Volatility Predictions in the Hong Kong Stock Market

0.99) except 12/97 (0.81). This is mainly due to the drastic reduction of the

contribution by the past conditional variance (ht-12). Notice that β1 drops from

0.91 to 0.37 within the period. Because of the severe drop of the HSI, despite

the sharp rise of the noise effect (α1) the drop of the β1 is even larger which

outweighs the noise effect. Thus, the sum of them is lowered.

Table 3(a) presents the parameters yielded by the Mixed model. As

before, α0 denotes the constant term, but now α1 denotes the partial elasticity

of the conditional variance with respect to its lagged value. For α1, its value

falls within a constant range (0.984 – 0.987), except during the AFC periods.

In 9/97, 12/97 and 3/98 it falls down to 0.981 or below. This indicates the

effect of external shock as random noise, which directly lowers the

autocorrelation of the conditional variances.

2. R2

After discussing the parameters, we then proceed to interpret the main

results. Table 4 presents the coefficient of determination (R2) by regressing the

HV on the average predicted volatility [h2Av(t,T)] of the RW, GARCH and

Mixed models respectively. Remember that only the set of 100–day evaluation

period (T=100) is used here.

39
A Comparison of Volatility Predictions in the Hong Kong Stock Market

In general the performance of the 3 models make no significant

differences: all of them perform poorly as the R2 is seldom above 0.7. Only in

9/96, 6/97, 3/98 and 9/98 do the 3 models yield a relatively high R2. This

seems rather occasional. On average the predictions are only good in 1/3 of the

time (or 4 out of 12), which may be better than wild guess but surely not

statistically acceptable. One important reason may be due to the problem of

the evaluation setting.

Observe the R2 cross-sectionally the problem seems not to be serious:

they are similar in the order of magnitudes, at least up to one decimal place in

general. From this, we conjecture that there may be some other problems like

the evaluation methodology – as it would be also awkward that all the 3

models are inappropriate but consistent to each other. At least the RW model

originates quite differently from the other two, where the former incorporates

no past information. As regression is commonly used in evaluation so we

present it here, but this does not mean it is a perfect way to do so.

As mentioned, regression requires quite a large sample to yield valid

result, after all, setting T = 30 is very marginal. On the other hand, the

volatility of the market has been really high in these years (and this enlightens

the objective of this paper). Whether making predictions over such a long

horizon as 100 trading days is reasonable (about the future 4 moths), is

certainly questionable. One may argue that T = 50 can be used instead.

40
A Comparison of Volatility Predictions in the Hong Kong Stock Market

However, remember that the average predicted volatility sets obtained will not

be accurate if the evaluation period is not long enough (i.e., T is not large

enough), even though the regression method is valid. That is, we have to

weigh 3 factors simultaneously in choosing a suitable T. Nevertheless, how T

should be chosen is purely the state of art, and there can be not much scientific

value in arguing the minor difference (e.g., between T = 80 or T = 90). A

better method will be presented in the next section.

But before doing so there are still “something” we can infer from

table 4. From the figure just below it, which is the plot of the data, it is clear

that all the 3 models have almost no predictive power in 9/97. This is

understandable as the crisis was so surprising that this observation should in

fact be discarded. However, the models did seem to do a good job in

predicting the period 3/98, 6/98 and 9/98. This can be explained as follows.

First, in the post-crisis period, the volume of the market contracted so

much that many people lost their assets or which were being tied up.

Especially when quite a proportion of them are expected to be speculators,

without their intense participation the volatility is expected to be smaller.

Second, the recession atmosphere worsened the market psychology. Even for

those who were able to invest would not be so brave in entering the market,

thus resulting in small volume and hence market contraction. Both of the

above reasons lead to the conclusion of small volatility. In general, the

41
A Comparison of Volatility Predictions in the Hong Kong Stock Market

predictability power of any regular model will be high under a data set with

relatively stable trend.

It seems odd in coming to the above conclusion. As mentioned earlier

that the volatility in the recent years have been high, how come a conclusion

that it was smaller?

The HSI plot given in the beginning of this paper may give us a

solution. Observe that in the post-crisis period, the index clustered around

10,000 to 12,000 from 12/97 to 4/98.16 From 6/98 to 10/98 it was around 6,600

to 8,600. That is, although in longer term the volatility was high, there were

certain shorter sub-periods where the index fluctuated only within some

ranges. And it is these 2 sub-periods that give rise to the high predictability in

3/98 and 9/98. On the contrary, a severe drop happened between the 2 sub-

periods (mainly from May to June) result in a relatively low predictability in

6/98 (R2 ≈ 0.6).17

Turn now to table 5, which shows the R2 by regressing the IV on the

average predicted volatility series of the 3 models respectively. It can be seen

that the results are even worse than that by using HV. None of them show an

acceptable level of R2: most of them are below 0.6. A very likely reason is that

the IV is not representative enough. Since the IV is obtained from only 6 of

the at-the-money series, its representative power is certainly questionable in

42
A Comparison of Volatility Predictions in the Hong Kong Stock Market

comparing with the HSI which consists of 33 blue chips. Nevertheless, due to

the limited size of the HSIO market, there is no way but to use this formal

published data given by the HKFE.

By comparing the results in tables 4 and 5 we find some consistency.

The major one is the extremely lowness of the R2 in 6/96, 9/97 and 12/98, with

the first two almost tend to zero. Ignoring the one in 6/96 where the

information period may not be long enough, in both 9/97 and 12/98

predictions there is one commonality – a large fluctuation of the HSI within

each of the period. The severe drop in 9/97 should be familiar, while the

strong rise of 3,000 points (from 7,500 to 10,500) happened between early

October and early November in 1998 (just in a month!). These 2 major

fluctuations occurred just within the evaluation periods with predictions made

at 9/97 and 12/98 respectively, thus much worsening the predictability. For

cross-sectional comparison, the conclusion is as before, none of the 3 models

significantly outperform the others.

3. MSD

Apart from evaluating the validity of the models which is shown to be

unsatisfactory by regression, another main purpose of this paper is to compare

the 3 models to see which is the best in prediction in the HK stock market

context. As the R2 obtained from regression “normalizes” the predictability of

the models to between zero and one, the difference between the models cannot

43
A Comparison of Volatility Predictions in the Hong Kong Stock Market

be significantly identified. Here then is the evaluation by using the mean

squared deviation (MSD) between the average predicted volatility and the

actual volatility (HV or IV). Unlike regression, there is no lower limit to the

sample size for validity consideration. So T can set to be any value. Table 6

shows the evaluation by using HV, while table 7 shows that by using IV. In

each set of tables, (a) to (d) represent T = 10 to T = 100 respectively. Notice

that all units are expressed in terms of 10-8.

Consider table 6 first. Unlike in the R2 section, for the MSD, it is the

smaller the better. Clearly the RW model is the worst in terms of predictability

except in few occasional cases. Then this amounts to compare the remaining 2

models: the GARCH and the Mixed. The following presents the count of the

number minimum MSD for each model in each table (i.e., for each T), where

“minimum” refers to the “least value” across models for each date (3/96, 6/96,

etc.). It follows that the more the count, the better the prediction.

Table 8(a): Count of minimum MSD with HV being used

Model \ T 10 25 50 100

GARCH 3 3 5 7

Mixed 9 8 6 4

It should be clear from the above table that for short-term prediction (T

= 10 or 25) the Mixed model does a very good job. But for long-term

44
A Comparison of Volatility Predictions in the Hong Kong Stock Market

prediction (T = 100) the GARCH model is the best. Viewed from the other

way, as T increases, the predictability of the GARCH model increases but that

for the Mixed decreases. Is this result true for using IV as actual volatility?

Table 8(b): Count of minimum MSD with IV being used

Model \ T 10 25 50 100

GARCH 4 5 5 6

Mixed 7 7 6 5

A similar result is found by using IV as actual volatility, but the

difference is not strong as that by using HV. Nevertheless, direct comparison

of the above count numbers across tables 8(a) and 8(b) is not meaningful for

the following reasons. First, the sample size is too small (only 12) that only

simple arithmetic analysis can be done on the above numbers. Second, the

standard for comparison (i.e., HV and IV) in the 2 tables are different in

underlying rationale. Thus, the difference across the 2 tables may not have

much theoretical implication.

Also, the RW model does the best in prediction occasionally, like 9/97

in table 6(b) and 12/97 in table 7(a). Moreover, the data (IV) for 12/98 in

tables 7(c) and (d) are not available yet in the research period, but this is not so

for those of HV in table 6. These explain why the vertical sum of the numbers

in either of the above 2 tables may not always be 12.

45
A Comparison of Volatility Predictions in the Hong Kong Stock Market

In spite of the above problems, the comparison among the difference of

the count numbers across the 2 models within the same table may still yield

some implications. That is, one conclusion is still reliable: for short-term

prediction the Mixed is the best model among, but for long-term prediction it

is the GARCH the best. It can also be inferred that there exists a (single or

region of) T ∈ (50,100) such that the predictions given by the 2 models are

“almost” indifferent. This is true regardless of which actual volatility (HV or

IV) is used for evaluation. We will see a more rigorous version of arriving this

claim in later paragraphs.

Before doing so, we may ask, “how is the above conclusion justified

theoretically?” This must deal with the models per se. Compare equation (5)

with (9), which are the GARCH and the Mixed models respectively. For the

moment ignore the random noise term in (9), the major difference between the

two is that the Mixed is in logarithmic form but the GARCH is not. Intuitively

whenever the volatility is in log form both its magnitude and its variance

would become drastically smaller. Think of a hypothetical situation where the

GARCH model gives a “perfect” prediction over the long horizon forecast. By

taking log to the variables (conditional variances and their lagged

components), one may well expect that it is the “log of the horizon” rather

than the “original horizon” would become the “perfect” prediction horizon.

46
A Comparison of Volatility Predictions in the Hong Kong Stock Market

Recall that after taking log the GARCH model is essentially the of

Mixed model. On the other hand, the log of the long horizon (the GARCH’s

“perfect” prediction horizon) is simply the short horizon – the Mixed’s

“perfect” prediction horizon. This argument is loose, but the aim is just to give

the picture of why the above conclusion is true in theoretical terms.

Even under the above conclusion, the MSD of the 2 models (GARCH

and Mixed) are of similar order of magnitude. This can be reasoned as they

belong to the same class of model (ARCH). Why does the RW model give

prediction with so large MSD? Observe that in general, for T increases from

10 to 100, the MSD changes from tenth to hundredth and then to thousandth,

in 10-8 unit. This is obvious from equation (3), where the average predicted

volatility over T days is obtained by multiplying T to the unconditional

variance. The RW model assumes the noise to be perfectly random, therefore

in predicting the average volatility in T day’s future T is multiplied to capture

the cumulating effect.

However, this is generally not the case for stock markets. Whenever

there is speculative element in the market, market psychology becomes

important in determining the trend, given that the market is not perfectly

efficient. In other words, the more investors’ concern on the past information,

the less powerful the RW model is. This can be evident from tables 6 and 7.

Observe that the MSD of the RW model is the lowest in 9/97 and 9/98, notice

47
A Comparison of Volatility Predictions in the Hong Kong Stock Market

that the first is just the post-crisis period, while the second is the rebound

period.

Of course, it cannot be said that people did not use information in those

periods when they participated in the stock market. Rather, information was

useless in prediction. In this way, the RW model is only accurate in the sense

of ex post “evaluation”. Unless one knows the crisis will happen beforehand,

there is no way to know when to use it. In spite of this, in the HK context, the

randomness of the market is far below the level where the RW model would

outperform the other two. This is evident from the comparatively large in

MSD values of the RW model relative to the other two in 9/97 and 9/98, even

though they are the smallest within the 12 (RW) model samples.

Having examined the minimum MSD, we turn now to the average of it.

The 12 predictions made in different dates are treated as the samples of a

particular model prediction with comparison to a particular actual volatility.

Instead of counting the number of minimum MSD across samples (i.e., dates)

as before, the average MSD is obtained for each model in each forecast

horizon (T). The results by using HV and IV as standard for comparisons are

presented in the following 2 tables:

48
A Comparison of Volatility Predictions in the Hong Kong Stock Market

Table 9(a): Average MSD with HV being used

T \ Model RW GARCH Mixed

10 19.35861 1.280413 1.172644

25 104.3457 6.91306 5.833455

50 468.0128 7.813349 6.491197

100 1976.563 6.00381 10.81391

Table 9(b): Average MSD with IV being used

T \ Model RW GARCH SV

10 18.01941 1.446924 1.015418

25 103.4345 3.253377 2.452684

50 467.35 3.008228 2.606145

100 1975.61 3.228776 9.338667

From the above we can see that the average MSD for the RW model

are very close across the 2 tables. This can be attributed to its large predicted

values so that when compared to the small actual values the difference

becomes very minor. For both models, the MSD under the HV being used is

generally higher than that of under the IV.18 That means the predicted values

are closer to the IV than the HV. This seems to suggest that using a “poor”

indicator (IV) instead of a better one (HV) would yield a “better” prediction.19

49
A Comparison of Volatility Predictions in the Hong Kong Stock Market

This bold conclusion may not be true. For sure IV has its advantage

over HV, where the former is based on the Black-Scholes’ theory, but no

theoretical basis is given for the latter in this context. However, we cannot tell

whether the above difference is statistically significant, nor can we say that the

Black-Scholes’ theory is the truth such that evaluation is precise. Hence, we

cannot be simply deduce from this that IV did a better evaluation in this

context, where, again, the sample size is too small.

Nevertheless, one thing is clear for the GARCH and the Mixed models.

As move along from T = 10 to T = 100, the average MSD with the HV being

used fluctuates more than that with the IV being used in general. This implies

that using IV yields relatively stable evaluation results, even though we cannot

tell which actual volatility is more suitable (accurate) in evaluation. Given the

generally highness of the R2 by using HV (compare tables 4 with 5) and the

smallness of the MSD by using IV [compare tables 9(a) with 9(b)], how

should we choose between them?

We must then back to theories at this point. Recall that R2 measures

how much the dependent variable is explained by the independent variable,

while MSD only measures how much one variable deviates from the other.

That is, for MSD the dependent-independent relationship plays no role and

hence the relationship is associative rather than explanatory. The question is,

should the actual volatility be explained by the volatility predictor?

50
A Comparison of Volatility Predictions in the Hong Kong Stock Market

The answer is yes. In terms of model prediction, it is the R2 that is

more appropriate “to predict”. Nevertheless, saying a model predicts

something is based on the assumption that the data follow the underlying

model used where the model assumptions should have been satisfied. In

practice, especially when the volatility changes so drastically, we have no way

to ensure which model do the data belong to. Except for the RW model which

data are assumed to be chaotic in nature, but has been proven to be empirically

inappropriate in our context.

In terms of making money, a good model is the one which does well in

“practical prediction” rather than in “theoretical prediction”. Doing well under

the specification by the theory is one thing, whether there is practical

significance is another matter. In other words, with the guarantee that a model

is true a priori (i.e., the data satisfy the model assumptions), the evaluation

results will automatically incorporate the causal nature. It follows that the two

kinds of instruments for evaluation, for which R2 demonstrates causality and

MSD demonstrates correlation, do not differ by much in theoretical terms.

If we go back to the results given in tables 9(a) and 9(b), we will

realize that this is really true. As similar to before, obtain the minimum

average MSD for each T. It is clear that for whichever actual volatility (HV or

IV) is used in prediction, for T = 10, 25 and 50 the Mixed model does better,

51
A Comparison of Volatility Predictions in the Hong Kong Stock Market

while for T = 100 it is the GARCH does better. This is exactly the conclusion

we have arrived before!

Previously the count of minimum MSD is used, but now it is the

average of them. It is the average value rather than simply the count that is

more accurate and appropriate, hence it is so-called a “more rigorous

approach”. Recall the dilemma in choosing between R2 and MSD, the

underlying, and ultimate objective is to see if HV or IV should be used in

evaluation. The above results resolve everything: both yield the same

conclusion, where either one can act as a double check for the other!

52
A Comparison of Volatility Predictions in the Hong Kong Stock Market

CONCLUSION

We can summarize the results of this empirical as follows: in Hong

Kong’s case, the Random Walk (RW) model is not useful; the GARCH model

is the best model for short term volatility prediction; and the combination of

the EGARCH and Stochastic Volatility (SV) models, i.e., the mixed model, is

the best model for long-term volatility prediction.

How can we determine the threshold level between the short-term and

the log-term? A plot of the tables 9(a) and 9(b) show that the threshold level T

is about 60 days for HV and just above 50 days for IV, roughly. It follows that

there should normally be about 20 non-trading days within 50 trading days

(trading to non-trading days ratio is about 5:2 in a typical week). This is

equivalent to 2 to 3 months in total.

Immediately there is a question: are the results reliable? In academic

literature much concern is placed on the choice of specific period for

analysis.20 However, a good model should be able to sustain the testing of any

arbitrary period, but not sensitive to some periods. That is, a good model

should be able to subject to the unexpected extreme high or low volatility.

Viewed in this way, although in our context the choice of period for

study (i.e., 1993 – 1998) is rather arbitrary in theoretical terms,21 the models

53
A Comparison of Volatility Predictions in the Hong Kong Stock Market

are obviously not able to survive under unexpected changes, such as the

sudden drop and rise of the HSI in 9/97 and 9/98 respectively. But this is

understandable, given the ex ante nature of the prediction, the model does not

take into account the forthcoming unexpected events.

Though, as the volatility has been high in these years, the predictions

beyond those unexpected events are generally acceptable. This can be evident

from the closeness of R2 to 0.7 for using HV in evaluation, and the pretty

lowness of MSD in general (except for the RW model).

It has also been suggested that the returns are serially correlated, where

this is not taken into account in this paper. In much literature it has been

shown many times that stock index returns follow the positive first-order

autocorrelation. The autocorrelation is positively related to number of stocks

included in the index and weight given to the thinly traded stocks, but

negatively related to interval over which returns are calculated.

Nevertheless, since only 33 stocks (out of several hundreds) are

included in the HSI; and those big blue chips (which are actively traded)

constitute a large proportion of the index.22 Also, the return interval used in

our context is minimized into daily basis, of which the interval could unlikely

to be further sub-divided within a day. Thus, none of these factors which

contribute to the autocorrelation are significant.

54
A Comparison of Volatility Predictions in the Hong Kong Stock Market

To eliminate this first-order autocorrelation which arises from the

micro-structural factors, the returns series are transformed as follows:

Rt = α + β Rt-1 + ut

The original daily return series (Rt) is regressed on its lagged series (Rt-1) so

that the autocorrelation effect is eliminated, and the residuals (ut) is the “true”

return series that should in fact be used in volatility prediction. However, by

doing so with our return data set, it is found that this kind of transformation is

in fact not very necessary.23

Another important consideration is that the predictions made are based

on the conditional variance at the last point of the data set (ht2). This is clearly

sensitive to the choice of when to end the data set at t, which can be improved

by increasing the number of information periods (where only 12 was used in

our context) at different t. However, by doing so the corresponding evaluation

periods will overlap heavily, thus making the evaluation of prediction at some

particular dates (or period of dates) impossible.

55
A Comparison of Volatility Predictions in the Hong Kong Stock Market

APPENDIX24

Derivation of equation (6)

From (1) and (5) with εt = ht vt , we have:

ht2+ k = α 0 + (α 1vt2+ k −1 + β 1 ) ht2+ k −1

By induction we have:

k −1 m k
ht2+ k = α 0 + α 0 ∑∏ (α1vt2+ k −n + β1 ) + ht2 ∏ (α1vt2+ k − n + β1 )
m =1 n =1 n =1

Given vt is i.i.d. distributed, it follows:

k −1
Et (ht2+ k ) = α 0 + α 0 ∑ (α 1 + β1 ) m + (α1 + β1 ) k −1 (α1vt2 + β1 )ht2
m =1

γ −γ k
Et (ht2+ k ) = α 0 + α 0 + γ k −1 (α 1vt2 + β 1 )ht2
1−γ

Substitute the above expression with γ = α1 + β1 into the following,

1 T
2
h Av (t , T ) = ∑
T k =1
Et (ht2+ k )

equation (6) results.

56
A Comparison of Volatility Predictions in the Hong Kong Stock Market

NOTE

1. The formula of the CAPM is given by: E(Ri) = RF + β [E(RM) – RF], where

R is return, i stands for individual, F for risk free, M for market.

2. The materials in this and next sections are mainly extracted from Heynen

and Kat (1994).

3. See Poterba and Summers (1986), and French, Schwert and Stambaugh

(1987).

4. See Schwert (1989). Since the indices in Hong Kong are published widely,

there should be no estimation error for their corresponding returns.

5. Using either does not matter: both give almost the same results (at least up

to three decimal places) for our purposes.

6. See Jacquier, Polson, and Rossi (1993).

7. See for example, Melino and Turnbull (1990).

8. See the HKFE homepage: <http://www.hkfe.com/>

9. Using our data set from 1993 to 1998, the volatilities in non-trading and

trading days are 0.000,328 and 0.000,384 respectively.

10. The monthly volumes are available from the HKFE homepage, see also

note 8.

11. See also note 2.

12. For date t being set at the end of 1998, T can only be 38 at maximum since

the data are only available up to the end of January 1999.

57
A Comparison of Volatility Predictions in the Hong Kong Stock Market

13. Assuming there are 250 trading days in a typical year, each quarter

consists of about 63 trading days. Notice that the non-trading days are

discarded.

14. Results for date t being set at the end of 12/1998 and 9/1998 have the

maximum T of 38 and 98 respectively, for same reason as note 12.

15. This exception is not easily explained in this context, but does no harm to

the general trend as a whole.

16. Notice that in 1/98 that the HSI dropped below the lower range limit.

17. Within the 2 months from the mid-April to mid-June, the HSI dropped by

about 4,000 points (from 11,400 to 7,500). This argument is not exact as it

may seem, since the prediction made in 6/98 should be compared with the

actual volatility over the 100 days starting from 1/7/98. Nevertheless, it

may be argued that the prediction based on the highly volatile data set in

the information period is equally inaccurate.

18. An exception to this is value of the GARCH model at T = 10.

19. The IV is poor relative to HV in the sense that regression on the former

yields a lower R2. This is due to the lack of representativeness of the IV as

a standard for comparison, which has been discussed in the previous

section.

20. See for example, Akgiray (1989), French, Schwert and Stambaugh (1986).

21. But this is not arbitrary in practical terms, as the data are only available in

the period chosen.

58
A Comparison of Volatility Predictions in the Hong Kong Stock Market

22. Like the Hongkong and Shanghai Banking Corporation (HSBC), it takes

up almost 1/4 weight of the HSI.

23. The R2 of the regression is just 0.0009, despite the p-value of the estimated

slope coefficient is 0.258.

24. See Heynen, Kemna, and Vorst (1994).

59
A Comparison of Volatility Predictions in the Hong Kong Stock Market

BIBLIOGRAPHY

Akgiray, V., “Conditional Heteroskedasticity in Time Series of Stock Returns:

Evidence and Forecasts”, Journal of Business, 62 (1989), pp. 55 – 80.

Bollerslev, T., “Generalized Autoregressive Conditional Heteroskedasticity”,

Journal of Econometrics, 31 (1986), pp. 307 – 327.

Chan, K. C., Cheng, L. T. W., and Fung, J. K. W., “Stock Market Seasonality

in China: A GARCH Analysis”, BRC Papers on China, Business Research

Centre, School of Business, Hong Kong Baptist College.

Cheng, L. T. W., Fung, J. K. W., and Chan, K. C., “The Impact of Index

Options Trading on the Constituent Stocks in Hong Kong”, BRC Papers on

Financial Derivatives and Investment Strategies, School of Business, Business

Research Centre, Hong Kong Baptist University.

Chou, R. Y., “Volatility Persistence and Stock Valuations: Some Empirical

Evidence Using GARCH”, Journal of Applied Econometrics, 3 (1988), pp.

279 – 294.

Dassios, A., “Asymptotic Approximations to Stochastic Variance Models”,

Discussion Paper, London School of Economics, 1992.

60
A Comparison of Volatility Predictions in the Hong Kong Stock Market

Enders, W., Applied econometric time series, New York: John Wiley & Sons,

1995.

Engle, R. F., “Autoregressive Conditional Heteroscedasticity with Estimates

of the Variance of United Kingdom Inflation”, Econometrica, 50 (1982), pp.

987 – 1007.

Engle, R. F., “Statistical Models for Financial Volatility”, Financial Analysts

Journal, Jan – Feb 1993, pp. 72 – 78.

Engle, R. F., and Mustafa, C., “Implied ARCH Models from Options Prices”,

Journal of Econometrics 52 (1992).

French, K. R., and Roll, R., “Stock Return Variances: The Arrival of

Information and the Reaction of Traders”, Journal of Financial Economics, 17

(1986), pp. 5 – 26.

French, K. R., Schwert, G. W., and Stambaugh, R. F., “Expected Stock Return

Variances: The Arrival of Information and the Reaction of Traders”, Journal

of Financial Economics, 17 (1986), pp. 3 – 29.

Greene, W. H., Econometric Analysis, New York : Macmillan, 1990.

61
A Comparison of Volatility Predictions in the Hong Kong Stock Market

Heynen, R. C., and Kat, H. M., (1994) “Volatility Prediction: A Comparison

of the Stochastic Volatility, GARCH (1,1), and EGARCH (1,1) Models”, in

the Journal of Derivatives, (1994), pp. 50 – 66.

Heynen, R. C., Kemna, A., and Vorst, T., “Analysis of the Term Structure of

Implied Volatility”, Journal of Financial and Quantitative Analysis, 29, 1

(1994), pp. 31 – 56.

Hong Kong Futures Exchange homepage.

Hull, J., and White, A., “The Pricing of Options on Assets with Stochastic

Volatilities”, Journal of Finance, 2 (1987), pp. 281 – 300.

Jacquier, E., Polson, N. G., and Rossi, P. E., “Bayesian Analysis of Stochastic

Volatility Models”, Working Paper, Graduate School of Business, University

of Chicago, 1993.

Kat, H. M., “Hedging Lookback and Asian Options”, Working Paper,

MeesPierson Derivatives, Amsterdam, 1993a.

Kat, H. M., “Replicating Ordinary Call Options: A Stochastic Simulation

Study”, 13th AMEX Options and Derivatives Colloquium, New York, 1993b.

62
A Comparison of Volatility Predictions in the Hong Kong Stock Market

Melino, A., and Turnbull, S. M., “Pricing Foreign Currency Options with

Stochastic Volatility”, Journal of Econometrics, 45 (1990), pp. 239 – 265.

Nelson, D. B., “ARCH Models as Diffusions Approximations”, Journal of

Econometrics, 45 (1990), pp. 7 – 38.

Nelson, D. B., “Conditional Heteroskedasticity in Asset Returns: A New

Approach”, Econometrica, 59 (1991), pp. 347 – 370.

Noh, J., Engle, R. E., and Kane, A., “Forecasting Volatility and Option Prices

of the S&P 500 Index”, The Journal of Derivatives, Fall 1994, pp. 17 – 30.

Poterba, J., and Summers, L., “The Persistence of Volatility and Stock Market

Fluctuations”, American Economic Review, 76 (1986), pp. 1142 – 1151.

Ruiz, E., “Quasi-Maximum Likelihood Estimation of Stochastic Volatility

Models”, Working Paper, London School of Economics, 1993.

Schwert, G. W., “Why Does Stock Market Volatility Change Over Time?”

Journal of Finance, 44 (1989), pp. 1115 – 1153.

63
A Comparison of Volatility Predictions in the Hong Kong Stock Market

Theodore, E. D., and Craig, M. L., “Stock Market Volatility and the

Information Content of Stock Index Options”, Journal of Econometrics, 52

(1992).

麥萃才, 期貨與期權, 香港: 中華, 1997.

64