This article was downloaded by: [The Aga Khan University]

On: 23 December 2014, At: 21:40

Publisher: Routledge
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer
House, 37-41 Mortimer Street, London W1T 3JH, UK

Applied Financial Economics

Publication details, including instructions for authors and subscription information:
http://www.tandfonline.com/loi/rafe20

Forecasting UK stock market volatility

a b c
David McMillan , Alan Speight & Owain Apgwilym
a
Department of Economics, University of St Andrews, Fife, KY16 9AL
b
Department of Economics, University of Wales, Swansea SA2 8PP, UK
c
Department of Management, University of Southampton, Southampton, SO171BJ
Published online: 07 Oct 2010.

To cite this article: David McMillan , Alan Speight & Owain Apgwilym (2000) Forecasting UK stock market volatility,
Applied Financial Economics, 10:4, 435-448, DOI: 10.1080/09603100050031561

To link to this article: http://dx.doi.org/10.1080/09603100050031561

PLEASE SCROLL DOWN FOR ARTICLE

Taylor & Francis makes every effort to ensure the accuracy of all the information (the “Content”)
contained in the publications on our platform. However, Taylor & Francis, our agents, and our licensors
make no representations or warranties whatsoever as to the accuracy, completeness, or suitability
for any purpose of the Content. Any opinions and views expressed in this publication are the opinions
and views of the authors, and are not the views of or endorsed by Taylor & Francis. The accuracy of
the Content should not be relied upon and should be independently verified with primary sources of
information. Taylor and Francis shall not be liable for any losses, actions, claims, proceedings, demands,
costs, expenses, damages, and other liabilities whatsoever or howsoever caused arising directly or
indirectly in connection with, in relation to or arising out of the use of the Content.

This article may be used for research, teaching, and private study purposes. Any substantial or
systematic reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution
in any form to anyone is expressly forbidden. Terms & Conditions of access and use can be found at
http://www.tandfonline.com/page/terms-and-conditions
 Applied Financial Economics, 2000, 10, 435 ± 448

Forecasting UK stock market volatility

D A V I D M C MI L L A N , A L A N S PE I G H T *{ and OW A I N A PG WI L Y M}
Department of Economics, University of St Andrews, Fife, KY 16 9AL { Department of
Economics, University of W ales, Swansea SA2 8PP, UK and } Department of Manage-
ment, University of Southampton, Southampton, SO17 1BJ

The paper analyses the forecasting performance of a variety of statistical and econo-
Downloaded by [The Aga Khan University] at 21:40 23 December 2014

metric models of UK FTA All Share and FTSE100 stock index volatility at the
monthly, weekly and daily frequencies under both symmetric and asymmetric loss
functions. Under symmetric loss, results suggest that the random walk model pro-
vides vastly superior monthly volatility forecasts, while random walk, moving aver-
age, and recursive smoothing models provide moderately superior weekly volatility
forecasts, and GARCH, moving average and exponential smoothing models provide
marginally superior daily volatility forecasts. If attention is restricted to one fore-
casting method for all frequencies, the most consistent forecasting performance is
provided by moving average and GARCH models. More generally, results suggest
that previous results reporting that the class of GARCH models provide relatively
poor volatility forecasts may not be robust at higher frequencies, failing to hold here
for the crash-adjusted FTSE100 index in particular.

I . I N T R OD U C T I ON using econometric techniques. Third, increased recognition

has been paid to the practical importance of accurate vola-
Since the stock market crash of 1987, stock price volatility tility estimates and forecasts in asset and option pricing
has been the focus of both empirical academic research and models, and portfolio selection and market timing deci-
regulatory concern. This attention re¯ ects three important sions (Gemmill, 1993; Vasilellis and Meade, 1996). These
considerations . First, the well-noted observation that stock considerations have led to examinations of the stationarity
market volatility has been higher in more recent than his- and persistence of volatility over time and the accuracy of
torical periods, and the perception that such increased volatility forecasting techniques. In particular, the com-
volatility is due to institutional changes such as automated parative forecasting performance of statistical procedures
trading and the introduction of trading in derivative associated with `technical analysis’ 1 , other more naõÈ ve
futures and options contracts which may have enhanced methods, and forecasts generated from extended models
the likelihood of large swings in mean stock returns of the GARCH class (Bollerslev et al., 1992).
(Edwards, 1988a, b; Schwert, 1990; Robinson, 1994; While the modelling and forecasting of US stock
Antoniou and Holmes, 1995). Second, although the ten- market conditional volatility has been extensively investi-
dency for stock market volatility to exhibit `clustering’ gated, the analysis of conditional volatility in other
has long been recognized (Mandelbrot, 1963; Fama, international stock markets has only recently been
1965), it is only since introduction of the autoregressive undertaken, with con¯ icting results. 2 For example, in fore-
conditional heteroscedasticity (ARCH) model by Engle casting monthly US stock market volatility, Akgiray (1989)
(1982) and its subsequent generalization (GARCH) by ® nds that a GARCH(1, 1) model outperforms more
Bollerslev (1986) that researchers have formally modelled traditional technical analysis. Tse (1991) and Tse and
the second and higher moments of ® nancial time series Tung (1992) examine the stock markets of Japan and

* Corresponding author. E-mail: [email protected]

1
For extended discussions of technical analysis and associated `chartism’, and the formal underpinnings of such analysis, see Stewart
(1966) and Plummer (1989) respectively.
2
Also see de Jong et al. (1992) , and Kearns and Pagan (1993).
Applied Financial Economics ISSN 0960± 3107 print/ISSN 1466± 4305 online # 2000 Taylor & Francis Ltd 435
http://www.tandf.co.uk/journals
 436 D. McMillan et al.
Singapore respectively, and ® nd that an exponentially GARCH model introduced by Engle and Lee (1993)
weighted moving average method is superior to the which permits the separation of volatility into long-term
GARCH model in both cases. Franses and van Dijk and short-term e€ ects and has not previously been consid-
(1996) compare the volatility forecasting performance of ered in the literature. In total, ten models are considered,
the GARCH, quadratic-GARCH (QGARCH, Engle and these being the historical mean, moving average, random
Ng, 1993), and threshold-GARCH (TGARCH, Glosten et walk, exponential smoothing, exponentially weighted mov-
al., 1992) models to the random walk model using weekly ing average, simple regression, GARCH, TGARCH,
German, Dutch, Italian, Spanish and Swedish stock index EGARCH, and component-GARCH models. Further, all
returns over the period 1986 to 1994. They report that the forecast evaluations are performed for the monthly, weekly
random walk model performs particularly well when the and daily data frequencies in order that the sensitivity of
crash of 1987 is included in the estimation sample, while our results to the choice of sample frequency may be
the QGARCH model performs better upon its exclusion. assessed. This exercise also has a practical motivation in
For the monthly volatility of Australian stock market data, that while tactical short-term investment decisions and
Brailsford and Fa€ (1996) compare the predictive valuations of short-lived derivatives frequently focus on
performance of several statistical methods with GARCH short-term predictions of volatility, strategic decisions
Downloaded by [The Aga Khan University] at 21:40 23 December 2014

and TGARCH models. Using several loss functions, they and the valuation of long-lived assets require longer-term
are unable to identify a clearly superior model and suggest predictions of volatility. The consideration of volatility
that the `best’ forecasting model depends upon the sub- forecasting performance over di€ erent frequencies and
sequent application, as the ranking of models is sensitive thereby di€ erent horizons, provides one means of accom-
to the choice of loss function. However, to date, little evi- modating such considerations and is an issue likely to be of
dence is available for the UK. Dimson and Marsh (1990) interest to such market participants as managers of index
examine various technical methods of predicting the vola- tracking funds who need to be aware of the volatility of
tility of UK stock market returns over the period 1955± these indices at di€ erent frequencies. All forecasts are also
1989 and ® nd that exponential smoothing and simple conducted on the basis of in-sample conditional mean
regression models perform best according to their criteria, models both with and without adjustment of the 1987
but do not consider the genre of ARCH models. Indeed, no crash. Finally, forecast evaluation is conducted with
analysis of the volatility forecasting performance of respect to a variety of loss functions. At a practical level
GARCH models for the UK stock market has yet been it is unlikely that all market participants will attach equal
reported. importance to overpredictions and underpredictions of
Several issues therefore arise from the existing empirical volatility, as in the pricing of call options on the buying
literature. First, there is no clear consensus across inter- and selling sides of the market for example, and both sym-
national stock markets as to the superior conditional volati- metric and asymmetric loss function results are therefore
lity forecasting method. Second, that these con¯ icting results reported.
in part re¯ ect di€ erences in the choice of classes of model for The remainder of the paper is structured as follows.
inclusion in the forecasting competition. Third, such results Section II describes the data series, source, sample period
are obtained for di€ erent frequencies of observation of stock and methodology. Section III describes the volatility fore-
market prices, and with di€ ering treatments of the 1987 casting models employed and characteristics of their in-
crash. Fourth, that the identi® cation of a superior volatility sample estimation. Section IV discusses the criteria on
forecasting method may not be invariant to the choice of loss which the forecast performance of these models is assessed
function used in forecast appraisal. and Section V reports the outcomes of the comparative
This paper therefore provides a comparative evaluation forecast exercise. Section VI summarizes our ® ndings and
of the volatility forecasting ability of GARCH models, concludes.
asymmetric TGARCH and exponential-GARCH
(EGARCH, Nelson, 1991) models, and a wide range of
more traditional methods associated with technical analy- II. DA TA
sis, for the Financial Times± Stock Exchange 100 index and
the Financial Times± Actuaries All Share index at the The Financial Times± Stock Exchange 100 index (FTSE100)
London Stock Exchange. A primary aim of the paper is was launched in January 1984 and represents the stock
to attempt to clarify the usefulness of GARCH type models prices of the 100 largest listed UK companies weighted by
in predicting volatility while using a di€ erent stock market market capitalization. One of the main aims of launching
setting from previous research, and to identify the `best’ the index was to enable the introduction of derivatives
forecasting models for the volatility of these indices of based on a single indicator of the equity market. It therefore
returns on the London stock exchange in particular. constitutes the ® rst real-time index for the UK equity
Additionally, and in extension of previous research inves- market which is small enough to be updated regularly and
tigating such models, we also examine the component- accurately but also comprehensive enough to act as an
 Forecasting UK stock market volatility 437
e€ ective proxy for the UK equity market, and is now a (out-of-sample) observations of 312 (19), 1357 (82) and
widely used indicator of UK equity market conditions. 3 6783 (413) for the monthly, weekly and daily FTA series,
The FTSE100 is thus a blue-chip subset of the Financial and 132 (20), 574 (83) and 2869 (413) for the monthly,
Times± Actuaries (FTA) All Share index introduced in weekly and daily FTSE series.
1962 which covers about 93% of the market capitalization In order to generate an historical `actual volatility’ series
of the UK equity market and is the long-standing bench- on the basis of which volatility forecasts may be generated
mark index for the UK market. While the FTSE100 and using statistical methods, we follow Pagan and Schwert
FTA All Share index are closely correlated, the di€ erent (1990) in representing true past volatility by the squared
composition of the two indices means there is always residuals from a conditional mean equation for returns
some scope for divergence between them and both are estimated over the in-sample period. Examination of the
therefore analysed here. in-sample returns data reveals signi® cant serial correlation
All data are obtained from the Datastream market infor- in the levels of the weekly and daily but not monthly
mation service. For the FTSE100 index we consider daily series. 4 The in-sample conditional mean is therefore esti-
closing price data from 2 January 1984 to 31 July 1996, mated using the autoregressive (AR) process :
while for the FTA All Share index we study daily closing ³’ …L †rt ˆ · ‡ ut ; ³’ …L † ˆ 1 ¡ ³1 L ¡ ¢ ¢ ¢ ¡ ³’ L ;
Downloaded by [The Aga Khan University] at 21:40 23 December 2014

price data from 1 January 1969 to 31 July 1996. These price

indices … p† are converted to returns …r† by the standard
t ˆ 1;2;. . . ;T …1†
method of calculating the log-di€ erence, rt ˆ
k
log … pt =pt¡1 †. Further, in order to examine the performance where L is the orthodox lag operator …L xt¡k ˆ xt¡k † and T
of the various volatility estimators over di€ erent forecast denotes the number of in-sample observations . 5 With esti-
horizons, the data are re-sampled at a weekly and monthly mates of the parameters of ³’ …L † and · determined by
returns frequency. Additionally, given the data period minimization of the in-sample sum of squared residuals of
extends over the crash of October 1987, volatility estima- Equation 1, historical volatility, ¼2t using the conventional
tion is conducted both including and excluding a dummy notation in the literature, is formally generated using:
variable for this event in order that the sensitivity of our 2
¼2t ˆ …rt ¡ f…ºT ¡ ‰ …³’;T …L † ¡ 1†rt Š†g† ;
results to this extreme volatility event may be assessed.
Finally, for each frequency the data are partitioned into
t ˆ 1;2;. . . ;T ‡ ½ …2†
the in-sample estimation periods 1969± 1994 for the FTA
data and 1984± 1994 for the FTSE data, the out-of-sample where ½ denotes the number of out-of-sample observations,
forecast periods covering the remaining period 1995± 1996 and the subscript T on a coe cient indicates that it is
for both data sources. This separation provides in-sample estimated conditional on the in-sample information set.6
3
The base level of the index was set at 1000. 0 on 31 December 1983, and it is now calculated on a ® rm basis by the London Stock
Exchange every 60 seconds during market hours, currently 0830± 1630 GMT.
4
For example, the Ljung± Box Q-statistics for up to twelfth-order autocorrelation for the FTSE (FTA) series are 36.43 (71.74) for weekly
returns and 47.04 (325.45) for daily returns.
5
Empirically, the lag length ’ is determined by the Akaike and Schwarz information criterion. Additionally, dummies for Monday and
holiday e€ ects are included in the AR model for daily data, while a January dummy is included for monthly data. Typically, low-order
AR models are determined. In particular, for the FTSE, an AR(0) model is appropriate for monthly data, AR(2) for weekly data, AR(1)
for daily data including a dummy for the 1987 crash, and AR(4) for daily data without a crash dummy. For the FTA series, an AR(2)
model holds for monthly data, an AR (3) for weekly data and AR(10) for daily data. The AR(10) in daily FTA data is a result of the two
week settlement-of-account period operating prior to reform in 1995. Thus an investor could have bought at the beginning of the
settlement period but held both cash and the asset, and returns thus compensate for this (see Theobald and Price, 1984). In 1995 this was
changed to a rolling settlement system.
6
Thus, historical volatility is generated by the application of a conditional mean model determined on in-sample data to both the in-
sample and out-of-sample returns data. While not discussed in the text, for several of the forecasting methods described in Section III, we
have also considered fully recursive variants of Equations 1 and 2 entailing re-estimation of the parameters …³’ …L †; ·† as the data sample
is extended over the out-of-sample horizon. Indeed, this approach is required in the recursive estimation and forecasting of the GARCH
class of models. However, this alternative, fully recursive, method of generating ¼2t imposes severe computational burdens in the context
of the moving average, smoothing and simple regression frameworks, since at each point in time over the out-of-sample forecast horizon
the entire history of volatility must then be reassessed and revised. Given that one objective of the present study is to ascertain the relative
bene® ts of alternative forecasting mechanisms in a practical context entailing repeated forecasting exercises, it is likely that simplicity is a
desirable feature of any recommended forecasting method, and the fully recursive approach may not be favoured under these methods on
such grounds. Moreover, the conditional mean functions estimated exhibit inherent stability. For example, comparison of the in-sample
conditional mean functions and full-sample conditional mean functions (the latter being employed in generating forecast error statistics,
see Section IV), reveal only minor di€ erences in coe cient estimates beyond the fourth decimal place. Thus, for those statistical methods
we have investigated using the fully recursive approach to the generation of historic volatility, whether that approach is employed or not
yields little or no quantitative di€ erence with the forecast error statistics reported in section V, and no qualitative di€ erence in terms of
relative forecasting performance across models.
 438 D. McMillan et al.
This notation is employed throughout the remainder of the data, two-and-one-hal f years and one-and-one-quarte r years
paper. for weekly data, and six months and three months for daily
data, corresponding to T ˆ 120 and T ˆ 60 data points for
the longer and shorter window lengths in each case. 7

I I I . V OL A T I L I T Y F OR EC A S T I N G M OD E L S
Random walk
Historical mean The preceding models presume reversion to a stable or
gradually shifting trend in volatility. However, if volatility
Extrapolation of the historical mean in volatility provides
¯ uctuates randomly the optimal forecast of next periods
perhaps the most obvious means of forecasting future vola-
tility. Moreover, if the distribution of volatility has a sta- volatility is simply current actual volatility:
tionary mean, all variation in estimated volatility is ht‡1 ˆ ¼2t ; t ˆ T ;T ‡ 1;. . . ;T ‡ ½ ¡ 1 …6†
attributabl e to measurement error and the historical
mean, ¼, - computed as the unweighted average of volatility This `random walk’ model thus suggests that the
observed in-sample, then gives the optimal forecast of vola- optimal forecast of volatility is for no change since the
Downloaded by [The Aga Khan University] at 21:40 23 December 2014

tility, h, for all future periods : last true observation. This model also provides us with
an alternative benchmark for appraising the relative fore-
T
1X casting performance of methods considered here, being a
ht‡1 ˆ ¼- 2 ˆ ¼2 ; t ˆ T ;T ‡ 1;. . . ;T ‡ ½ ¡ 1
T jˆ1 j standard comparative method in econometric forecast
appraisal.
…3†
Forecasts based on this mean also provide a benchmark for
the comparative evaluation of the alternative forecasting Exponential smoothing
models outlined below. In addition to this in-sample histor- Under exponential smoothing the one-step-ahead
ical mean, we also consider the recursive assessment of the forecast of volatility is a weighted function of the
historical mean, iteratively updated with each incremental immediately preceding volatility forecast and current
observation on volatility over the out-of-sample period: actual volatility:
t
X
1 ht‡1 ˆ ¿T ht ‡ …1 ¡ ¿T †¼2t ; t ˆ T ;T ‡ 1;. . . ;T ‡ ½ ¡ 1
ht‡1 ˆ ¼- 2t ˆ ¼2j ; t ˆ T ;T ‡ 1;. . . ;T ‡ ½ ¡ 1 …4†
t jˆ1 …7†
such that the mean of historic volatility and forecast of where the `smoothing parameter’ is constrained such that
future volatility at any point in time during the out-of- 0 µ ¿ µ 1. For ¿ ˆ 0 (or ht ˆ ¼2t , an exactly correct past
sample period is based on all information on actual vola- forecast) the exponential smoothing model collapses to the
tility available at that point in time. random walk, while as ¿ ! 1 major weight is given to the
prior period forecast. As the subscripts indicate, the value
Moving average of ¿T is determined empirically as that which minimizes the
in-sample sum of squared prediction errors. 8
Under the moving average method volatility is forecast
by an unweighted average of past observed volatilities
over a particular historical time interval of ® xed Exponentially weighted moving average (EW MA)
length : The exponentially weighted moving average model is
1 t
X similar to the exponential smoothing model, but where
ht‡1 ˆ ¼- 2t;T ˆ ¼2 ; t ˆ T ; T ‡ 1;. . . ;T ‡ ½ ¡ 1 past observed volatility in Equation 7 is replaced with a
T jˆT ¡T j
moving average forecast as in Equation 5:
…5† t
1 X
where T is the moving average period or `rolling window’. ht‡1 ˆ ÁT ht ‡ …1 ¡ ÁT † ¼2 ;
T jˆT ¡T j
The choice of this interval is essentially arbitrary and two
lengths are considered here for each frequency. These
arbitrary choices are ten years and ® ve years for monthly t ˆ T ;T ‡ 1;. . . ;T ‡ ½ ¡ 1 …8†
7
Thus, approximately the same number of observations are used in constructing the moving averages over the three sample frequencies.
8
The values estimated here lie in the range 0.001± 0.1 for monthly and weekly data, and 0.14± 0.83 for daily data. This contrasts with
results reported elsewhere. For example, Dimson and Marsh (1990) using quarterly volatilities report a value of 0.66 in their preferred
model, while Brailsford and Fa€ (1996 ) report values in the range 0.51 to 0.98 for monthly data.
 Forecasting UK stock market volatility 439
with T speci® ed as for the longer of two horizons considered quantifying volatility, out-of-sample forecasts of volatility
for each frequency in the moving average model above. 9 are generated by the GARCH…p; q† model:12
F 0
’;T …L †rt ˆ ·T0 ‡ "t;
Simple (mean) regression
The simple regression model provides one-step-ahead fore- F 0
’;T …L † ˆ 1 ¡ F 0
1;T L ¡ ¢¢¢ ¡ F ’;T L ; t ˆ 1;2;. . . ;T ‡ ½ …10†
casts generated from the application of an in-sample esti-
mated ordinary least squares regression of observed actual "t ¹ N…0;ht † …11†
volatility upon immediately preceding actual volatility to
out-of-sample data: q
X p
X
ht‡1 ˆ !T ‡ ¬i ;T "2 t¡i‡1 ‡ ­ i;T ht¡i‡1 ;
ht‡1 ˆ ® T ‡ ¯T ¼2t¡1 ; t ˆ T ;T ‡ 1;. . . ;T ‡ ½ ¡ 1 …9† iˆ1 iˆ1

Following Dimson and Marsh (1990), again assuming the t ˆ T ;T ‡ 1;. . . ;T ‡ ½ ¡ 1 …12†
stationarity of volatility over the longer term, if such P P
where ! > 0, ¬i , ­ i ¶ 0, and i ¬i ‡ i ­ i < 1, the latter
forecasts are to be unbiased then the simple regression
sum quantifying the persistence of shocks to volatility. 13
Downloaded by [The Aga Khan University] at 21:40 23 December 2014

implicitly forecasts volatility as a weighted sum of recent

The GARCH(1, 1) model, for example, thus generates one-
volatility and long-run mean volatility, and such that
step-ahead forecasts of volatility, ht‡1 , as a weighted average
volatility will regress from its most recent level, ¼2t¡1 ,
- of the constant long-run or average variance, !, the pre-
towards its long-run mean h, with ¯ determining the
viously forecast variance for the current period, ht , and cur-
speed of regression towards that mean. 10 ;11
rent volatility re¯ ecting squared `news’ about the index
return, "2t . In particular, volatility forecasts are increased
Generalized autoregressive conditional heteroscedasticity following a large positive or negative index return, the
(GARCH) GARCH speci® cation thus capturing the well-documented
volatility clustering evident in ® nancial returns data. 14
The GARCH model of Engle (1982) and Bollerslev (1986)
requires joint estimation of the conditional mean and vari-
ance process, the former being represented by an auto-
T hreshold-GARCH (T GARCH)
regressive process for stock index returns as above in
Equation 1. On the assumption that the resulting con- The GARCH model, although non-linear in the con-
ditional mean stochastic error, "t , is normally distributed ditional mean error term, "t , postulates a linear relation-
with zero mean and time-varying conditional variance ship between forecast volatility, previous forecasts of
9
As the subscript notation indicates, the value of Á is again determined empirically by minimization of the in-sample residual sum
of squares and found to be 0.999 for both the FTSE and FTA series at all three frequencies, with the exception of the monthly-
sampled FTSE where the value is 0.4340 for data adjusted for the 1987 stock market crash and 0.5060 without such adjustment. This
compares with a range of values reported by Brailsford and Fa€ (1996) of between 0.0 and 0.9. Additionally, it should be noted that
models were also estimated for the shorter moving average intervals discussed, but with no qualitative di€ erence from the results
reported here. -
10
To see this, note that the expected
- value of both sides of Equation
- 9 must equal the long-run mean, h, for unbiasedness to hold. Hence,
2
® may be replaced by …1 ¡ ¯†h …by virtue of E …¯¼t¡1† ˆ ¯h†, yielding:
-
ht‡1 ˆ …1 ¡ ¯T †h ‡ ¯T ¼2t¡1 ; t ˆ T ; T ‡ 1; . . . ; T ‡ ½ ¡ 1 …9b†
11
In-sample estimates of ¯ are typically low and positive in the lower frequency data, at 0.05 (0.11) and 0.08 (0.12) in monthly and weekly
FTA data (crash adjusted values in brackets), and 0.04 (70.06) and 0.01 (0.03) in monthly and weekly FTSE data. Only for the adjusted
monthly FTA and unadjusted weekly FTSE series are these values statistically signi® cantly di€ erent from zero. At the higher daily
frequency all estimated ¯ coe cients are positive and statistically signi® cant with values of 0.46 (0.35) and 0.63 (0.27) for the FTA and
FTSE respectively.
12
Note that in the GARCH model, "2t is the analogue of historic volatility, ¼2t , in the statistical forecasting models described above, the
only essential di€ erence being that while ¼2t is generated as the squared residuals from the conditional mean function estimated alone, "2t
is generated as the squared residuals from the conditional mean function estimated jointly with the conditional variance process in
Equations 10± 12. While alternative notation could be employed in clari® cation of this connection, we retain the representation employ-
ing "2t in keeping with conventions in the GARCH literature.
13
As Nelson and Cao (1992) have shown, the ® rst two inequalities stated need not necessarily hold in order to ensure non-negativity of
the conditional variance. In the
P GARCH(1,
P 2) case for example, ¬2 < 0 with ­ 1 ¬1 ¶ ¬2 is su cient to ensure h2t > 0. However, it is
necessary and su cient that i ¬i ‡ i ­ i < 1 in order P for a ® nite
P unconditional variance to exist, that sum also de® ning the limiting
case of the integrated-GARCH (IGARCH) model for i ¬i ‡ i ­ i ˆ 1, such that current news has a permanent e€ ect on forecast
volatility (Engle and Bollerslev, 1986).
14
Preliminary tests for the presence of ARCH e€ ects in in-sample conditional mean errors using Engle’s (1982) LM test indicate
signi® cant conditional heteroscedasticity in all cases other than the weekly FTSE series unadjusted for the 1987 crash, that event
 440 D. McMillan et al.
q
volatility, and current and lagged measured volatility in X
response to news. However, it has been observed that posi-
tive and negative shocks of equal magnitude have a di€ er-
log …ht‡1† ˆ !T ‡
iˆ1
… j"t¡i‡1 j
¬i;T ±1;T p 



j ht¡i‡1 j

" ‡1
‡ ±2;T pt¡i




†



ht¡i‡1


p
ential impact upon stock price volatility, which may be X
attributabl e to a `leverage’ e€ ect (Black, 1976). Further, ‡ ­ i;T log …ht¡i‡1† …14†
iˆ1
stock market returns series have been noted to display sig-
ni® cant negative skewness, possibly due to the fact that t ˆ T ; . . . ;T ‡ ½ ¡ 1, where the ±1 coe cient captures the
market crashes are greater in absolute size and occur volatility clustering e€ ect noted above and the coe cient ±2
more frequently and more quickly than booms (Franses measures the asymmetric e€ ect, which if negative indicates
and van Dijk, 1996). One model that is able to capture that negative shocks have a greater impact upon con-
these features is the TGARCH model (Glosten et al., ditional volatility than positive shocks of equal magnitude.
1993) which, for a ® rst-order threshold, is expressed as Additionally, the use of logs allows the parameters ¬i and
(in conjunction with Equations 10 and 11) : ­ i to be negative without the conditional variance becom-
ing negative, while the Ppersistence of shocks to conditional
q
X p
X variance is given by i ­ i . 16
ht‡1 ˆ !T ‡ ¬i ;T "2t¡i‡1 ‡ ¹T It "2t ‡ ­ i;T ht¡i‡1 ;
Downloaded by [The Aga Khan University] at 21:40 23 December 2014

iˆ1 iˆ1
Component-GARCH (CGARCH)
t ˆ T ;T ‡ 1;. . . ;T ‡ ½ ¡ 1 …13†
The CGARCH model of Engle and Lee (1993) attempts to
separate long-run and short-run volatility e€ ects in a simi-
where It ˆ 1 if "t < 0, and It ˆ 0 if "t > 0. Thus, in the
lar manner to the Beveridge± Nelson (1981) decomposition
TGARCH(1, 1) case, for example, positive `news’ has an
of conditional mean ARMA models for economic time
impact of ¬1 on volatility while negative news has an
series. Thus, while the GARCH model and its asymmetric
impact of ¬1 ‡ ¹, positive (negative) news therefore having
TGARCH and EGARCH extensions exhibit mean rever-
the greater impact on subsequent volatility for ¹ < 0
sion to !, the CGARCH model allows mean reversion to a
¹ > 0†, Pwhile shock persistence is quanti® ed by
…P time-varying level, !t . The CGARCH speci® cation is:
i ¬i ‡
= 15
i ­ i ‡ …¹ 2†.
ht‡1 ˆ !t‡1 ‡ ¬1 …"2t ¡ !t † ‡ ­ 1 …ht ¡ !t †;
t ˆ T ;T ‡ 1;. . . ;T ‡ ½ ¡ 1 …15†
Exponential-GARCH (EGARCH)
!t‡1 ˆ ! ‡ »!t ‡ ¹…"2t ¡ ht †
The EGARCH model (Nelson, 1991) provides an alterna-
tive asymmetric model, the conditional variance being where !t represents a time-varying trend or permanent
expressed as an asymmetric function of past errors (again component in volatility which is integrated if » ˆ 1. The
in conjunction with Equations 10 and 11) : volatility prediction error …"2t¡1 ¡ ht¡1† serves as the driving

swamping all other conditional variance behaviour in that series. For all remaining series, in-sample AR-GARCH model orders …’;p; q†
are initially determined by application of the Akaike and Schwarz information criterion. However, as noted by Nelson (1991, p. 356), the
asymptotic properties of the Schwarz criterion in the context of ARCH are unknown, while the Akaike criterion tends to prefer long lag
models. Thus, examination of the residual diagnostics and maximization of the log-likelihood are also used to guide order selection.
Estimation of GARCH models and extended GARCH models is performed jointly with the AR model throughout, and dummies for
possible Monday and holiday e€ ects are again included in the AR model for daily data, while a January dummy is again included for
monthly data. The resulting jointly estimated AR models orders remain the same as reported above under AR estimation alone, while
joint GARCH(1, 1) models hold for all series other than FTA daily data where a GARCH(3, 1) model holds. Persistence measures for
these models applied to FTA and FTSE data are found to be 0.90± 0.93 for monthly data, 0.93± 0.94 for weekly data and 0.95± 0.97 for
daily data, with the exception of the monthly FTSE data unadjusted for the 1987 crash where a persistence measure of 1.00 identi® es an
IGARCH process. Thus, GARCH volatility forecast mean reversion is absent for the latter series, and generally sluggish for the
remaining series in view of their relatively high shock persistence measures.
15
Estimation and forecasting of the TGARCH model is conducted along the same lines as for the GARCH model, the AR-GARCH
model orders reported above continuing to hold for jointly estimated AR-TGARCH models with ® rst-order thresholds imposed. The
estimated asymmetric coe cient, ¹, is everywhere positive and statistically signi® cant in-sample for the 1987 crash-adjusted FTA series at
all frequencies, the non-adjusted daily FTA series, the non-adjusted monthly FTSE series and the crash-adjusted weekly FTSE series.
Persistence measures remain virtually unchanged relative to those for in-sample GARCH models, except for the IGARCH crash-
unadjusted monthly FTSE case where persistence falls to 0.50, and the crash-unadjusted monthly FTA case where persistence falls to
0.84.
16
Estimation and forecasting of the EGARCH model is conducted as for GARCH and TGARCH models, model orders continuing to
hold for jointly estimated AR-EGARCH models. The asymmetric coe cient, ±2 , is consistently negative and statistically signi® cant in-
sample for the 1987 crash-adjusted FTA series at all frequencies, the non-adjusted FTA series at the daily frequency, and the FTSE
crash-adjusted series at both the monthly and weekly frequencies. Persistence measures con® rm those obtained for GARCH and
TGARCH models, with the exception of the monthly FTSE series where the previously integrated crash-unadjusted GARCH case
 Forecasting UK stock market volatility 441
force for the time-dependent movement of the trend, and error (RMSE) and mean absolute error (MAE), de® ned as
the di€ erence between the conditional variance and its follows:
trend …ht ¡ qt† de® nes the transitory component of the con-
ditional variance. The transitory component then con- 1 TX‡½

verges to zero with powers of …¬ ‡ ­ † while the long-run

ME ˆ …ht ¡ s2t † …17†
½ tˆT ‡1
component converges on ! with powers of ». 17
1 TX‡½
MAE ˆ jht ¡ s2t j …18†
Recursively estimated models ½ tˆT ‡1
In addition to recursive appraisal of the historical mean v
u










 

u 1 TX‡½
(see Equation 4 above) we also consider forecasts generated t
RMSE ˆ …ht ¡ s2t † …19†
by recursive variants of the above models which involve ½ tˆT ‡1
parameters estimated using in-sample data. That is, those
models involving parameters with a T subscript. These where ½ is the number of forecast data points and s2t is
alternative recursively generated one-step-ahead forecasts `true’ volatility as de® ned above. The ME statistic is used
Downloaded by [The Aga Khan University] at 21:40 23 December 2014

are based on re-estimation of the underlying parameters at here as a general guide to the direction of over- or under-
each data point over the out-of-sample period, that re-esti- prediction on average. The MAE is an orthodox forecast
mation utilizing all information available at that point in appraisal criterion which does not permit the o€ setting
time. Thus, while not speci® ed in full, the set of parameters e€ ects of overprediction and underprediction as in ME,
«T ˆ …¿T ;ÁT ;® T ;¯T ; F ’;T ; ·T ;!T ;¬i;T ; ­ i;T ;¹T ;±i;T ; while the RMSE is a conventional criterion which clearly
weights greater forecast errors more heavily in the average
±2;T ;»T † forecast error penalty.
These error statistics assume the underlying loss function
in Equations 7± 10 and 12± 15 are correspondingly replaced
to be symmetric. However, as noted in the introduction, it
by the set «t ˆ …¿t ;Át ;. . . ;»t † for t ˆ T ;T ‡ 1;. . . ;
is probable that as a practical matter not all investors will
T ‡ ½ ¡ 1. 18 The criteria on which the forecast perform-
attach equal weight to similar sized overpredictions and
ance of the resulting and the foregoing models are assessed
underpredictions of volatility. 19 Following previous
is discussed in the following section.
research (Pagan and Schwert, 1990; Brailsford and Fa€ ,
1996 ) we therefore also consider error statistics designed
to account for potential asymmetry in the loss function.
I V . F OR E C A S T E V A L U A T I ON That is, mean mixed error statistics which penalize, ® rstly,
underpredictions more heavily:
In order to provide a measure of `true’ volatility against "
O U q




#
which to assess the forecast performance of the volatility 1 X X
MME …U† ˆ jht ¡ s2t j ‡ jht ¡ s2t j …20†
estimators, we follow Pagan and Schwert (1990) in using ½ iˆ1 iˆ1
the squared error term from a conditional mean model for
returns estimated over the full data set comprising both the and secondly, overpredictions more heavily:
in-sample and out-of-sample data. That is, `true’ volatility " #
O q 





 X U
generated by: 1 X 2 2
MME …O † ˆ jht ¡ st j ‡ jht ¡ st j …18†
½ iˆ1
s2t ˆ …rt ¡ fº½ ¡ ‰ …³’;½ …L † ¡ 1†rt Šg†2 …16† iˆ1

where the subscript ½ on a coe cient indicates that it is where O denotes the number of overpredictions and U the
estimated over the entire data sample. The ability of the number of underpredictions among the out-of-sample fore-
above models to adequately forecast true volatility so casts. Finally, and again following previous research, we
measured in the FTSE and FTA stock index markets is also report standardized values for all error statistics using
evaluated using the mean error (ME), root mean squared the error statistic for the historical mean benchmark for

now obtains a persistence measure of ¡0:94, suggesting oscillation in conditional variance, and the corresponding crash-adjusted
measure falls to 0.43.
17
In-sample estimates of » generally con® rm earlier persistence measures, at 0.83 (0.64), 0.92 (0.91), and 0.99 (0.99) in crash-adjusted
(unadjusted) FTA monthly, weekly and daily series respectively, and 0.91 (0.68), 0.94 and 0.94 (0.92) in corresponding FTSE adjusted
(unadjusted) series. Persistence in the transitory component, ¬ ‡ ­ , is everywhere lower than in the permanent component, though only
marginally for daily FTA data, and negative for the monthly crash-adjusted and weekly unadjusted FTA series and adjusted daily FTSE
series, suggesting the oscillation of volatility about the long-run trend during convergence in those cases.
18
Note that, as mentioned in footnote 6, for the class of GARCH models such recursive estimation is `fully recursive’ in the sense that
not only are the volatility forecasting equation parameters iteratively updated with the acquisition of out-of-sample information, but the
 442 D. McMillan et al.
each series. This has the advantage of allowing the error ing average for FTSE data. Nevertheless, on the basis of
statistics to be more easily interpreted in a relative context. the MAE forecast error statistic, the random walk model
continues to dominate the forecasting performance for
both indices, followed by the exponential smoothing
model for the FTA series while the moving average and
V . F OR E C A S T R E S U L T S
recursive EWMA models provide the next best forecasting
performance on broadly similar forecast error statistics for
Symmetric forecast error results
both sets of FTA and FTSE data, followed by the GARCH
Tables 1A and 1B report forecast ME, RMSE and MAE and remaining smoothing models with broadly similar
statistics for the FTSE and FTA indices sampled at the forecast error statistics. However, under the RMSE
monthly frequency, unadjusted and adjusted for the 1987 forecast error statistic, both moving average models and
crash respectively. The ME statistic indicates that all the recursive exponential smoothing and exponential
models overpredict volatility for both series, with the sole EWMA models now provide equivalently superior fore-
exception of the random walk, which also provides the casts for the FTA series, both including and excluding
forecast with the smallest absolute ME for both series. the crash period. Similarly, both moving average models
Downloaded by [The Aga Khan University] at 21:40 23 December 2014

More substantively, the random walk model also domi- and the recursive EWMA model provide equivalently
nates on both the RMSE and MAE statistics for both superior forecasts for crash-unadjuste d FTSE data, while
series, and both including and excluding the crash period. the 1.25 year moving average is singularly superior for the
Additionally, as indicated by the standardized statistics, the crash-adjusted FTSE. On these criteria the GARCH mod-
gain in performance of the random walk model over all els perform comparably with the remaining smoothing
other models is considerable. The simple regression and methods, and more favourably in some instances, notably
historical mean perform generally poorly, particularly in the recursive GARCH and EGARCH models. Finally, on
forecasting FTA monthly volatility, providing the worst both the MAE and RMSE criteria, the simple regression
forecasts for that series. Among those other models, the and historical mean models again provide the poorest
performance of the GARCH and smoothing models forecasts for the FTA series, while the exponential smooth-
(moving average, exponential smoothing, and EWMA) is ing model provides the poorest forecasts for the FTSE
similar across measures and series, with some notable series, although the historical mean and simple regression
exceptions: the good performance of recursive exponential models again perform poorly.
smoothing for the FTA series; the poor performance of the Tables 3A and 3B report forecast error statistics for returns
non-recursive GARCH and TGARCH for the FTSE volatility on daily sampled unadjusted and crash-adjusted
series, particularly evident where poorer relative to the his- FTSE and FTA data. ME statistics indicate overprediction
torical mean benchmark; likewise the poor performance of for all models except the random walk and recursive expo-
the exponential smoothing models for the crash-unadjuste d nential smoothing in crash-adjusted series and, additionally,
FTSE series. Further, the ® rst of these exceptions apart, the (non-recursive ) exponential smoothing in the crash-
® ve-year moving average (and equally the recursive unadjusted series. Recursive exponential smoothing also
EWMA model for FTSE data) consistently marginally out- now gives the minimum ME statistic across all series other
performs other models. 20 than the crash-adjusted FTA, where the 3 month moving
Tables 2A and 2B present ME, RMSE and MAE statis- average model is preferred. For the FTA series, the exponen-
tics for forecasts of weekly sampled FTA and FTSE tial smoothing and three-month moving average models pro-
returns volatility, again crash-unadjuste d and crash- vide the best forecasts on the MAE statistic for crash-
adjusted respectively. As for the monthly frequency, all unadjusted and crash-adjusted data respectively, followed
models again generate overpredictions of volatility with by the remaining smoothing models (excepting the non-recur-
the exception of the random walk, although minimum sive EWMA) and GARCH models thereafter. This ordering
ME statistics are now obtained by the recursive exponen- also holds for crash-unadjusted FTSE data on the MAE cri-
tial smoothing model for FTA data and the 1.25 year mov- terion, but not the crash-adjusted FTSE series, for which the
conditional mean function for stock index returns, which is used to generate historic volatility (i.e. "2t in the GARCH context), is also
iteratively updated. On the implementation of similar procedures for the class of statistical forecasting models, see footnote 6.
19
For example, consider the positive relationship between the volatility of underlying stock prices and call option prices previously noted
by Brailsford and Fa€ (1996) . An underprediction of volatility will impart a downward bias to estimates of the call option price, which is
likely to be of more concern to a seller than a buyer, while the reverse is true of overpredictions of stock price volatility.
20
As noted in the Introduction, Dimson and Marsh (1990) reject the random walk model for quarterly volatility forecasting of the FTA,
preferring the simple regression and exponential smoothing models. One possible explanation for the divergence between their results
and ours is that Dimson and Marsh compute actual volatility as annualized standard deviations of daily returns over non-overlapping
calendar quarters, and generate one-step-ahead forecasts of that quarterly series. Nevertheless, their ® ndings receive some support here at
the monthly frequency in terms of our recursive exponential smoothing model results for FTA data, which prove `second best’ to the
random walk model.
 Forecasting UK stock market volatility 443
Table 1. Forecast error statistics; monthly frequency

FTA-All FTSE

Model ME MAE RMSE ME MAE RMSE

A. Excluding dummy for 1987 crash

Historical mean 0.0035 (1.00) 0.0035 (1.00) 0.0035 (1.00) 0.0025 (1.00) 0.0026 (1.00) 0.0026 (1.00)
Recursive historical mean 0.0034 (0.97) 0.0034 (0.97) 0.0034 (0.97) 0.0024 (0.96) 0.0024 (0.92) 0.0024 (0.92)
Moving average ± 5 years 0.0014 (0.40) 0.0012 (0.34) 0.0013 (0.37) 0.0013 (0.55) 0.0013 (0.50) 0.0014 (0.54)
Moving average ± 10 years 0.0026 (0.74) 0.0015 (0.43) 0.0015 (0.43) 0.0025 (1.03) 0.0025 (0.96) 0.0026 (1.00)
Random walk 70.0005 (0.13)* 0.0004 (0.11)* 0.0006 (0.17)* 70.0004 (0.16)* 0.0004 (0.15)* 0.0005 (0.19)*
Exponential smoothing 0.0016 (0.46) 0.0016 (0.46) 0.0017 (0.49) 0.0032 (1.30) 0.0032 (1.23) 0.0032 (1.23)
Recursive exponential smoothing 0.0009 (0.26) 0.0010 (0.29) 0.0011 (0.31) 0.0030 (1.20) 0.0030 (1.15) 0.0030 (1.15)
EWMA 0.0027 (0.76) 0.0019 (0.54) 0.0019 (0.54) 0.0017 (0.71) 0.0026 (1.00) 0.0026 (1.00)
Recursive EWMA 0.0026 (0.74) 0.0026 (0.74) 0.0026 (0.74) 0.0014 (0.56) 0.0014 (0.54) 0.0014 (0.54)
Simple regression 0.0033 (0.96) 0.0033 (0.94) 0.0034 (0.97) 0.0025 (1.00) 0.0025 (0.96) 0.0025 (0.96)
Recursive simple regression 0.0027 (0.76) 0.0027 (0.76) 0.0027 (0.76) 0.0023 (0.92) 0.0023 (0.88) 0.0023 (0.88)
GARCH 0.0020 (0.56) 0.0020 (0.57) 0.0020 (0.57) 0.0021 (0.84) 0.0021 (0.81) 0.0021 (0.81)
Downloaded by [The Aga Khan University] at 21:40 23 December 2014

Recursive GARCH 0.0018 (0.51) 0.0018 (0.51) 0.0019 (0.54) 0.0017 (0.68) 0.0017 (0.65) 0.0019 (0.73)
TGARCH 0.0022 (0.62) 0.0022 (0.63) 0.0022 (0.63) 0.0030 (1.22) 0.0028 (1.08) 0.0030 (1.15)
Recursive TGARCH 0.0021 (0.60) 0.0021 (0.60) 0.0021 (0.60) 0.0027 (1.08) 0.0027 (1.04) 0.0027 (1.04)
EGARCH 0.0016 (0.47) 0.0016 (0.46) 0.0018 (0.51) 0.0024 (0.96) 0.0024 (0.92) 0.0026 (1.00)
Recursive EGARCH 0.0015 (0.43) 0.0015 (0.43) 0.0017 (0.49) 0.0022 (0.88) 0.0022 (0.85) 0.0024 (0.92)
C-GARCH 0.0022 (0.62) 0.0022 (0.63) 0.0022 (0.63) 0.0028 (1.15) 0.0028 (1.08) 0.0029 (1.12)
Recursive C-GARCH 0.0021 (0.60) 0.0021 (0.60) 0.0022 (0.63) 0.0025 (1.00) 0.0026 (1.00) 0.0026 (1.00)

B. Including dummy for 1987 crash

Historical mean 0.0031 (1.00) 0.0031 (1.00) 0.0031 (1.00) 0.0018 (1.00) 0.0018 (1.00) 0.0018 (1.00)
Recursive historical mean 0.0030 (0.97) 0.0030 (0.97) 0.0030 (0.97) 0.0016 (0.89) 0.0016 (0.89) 0.0017 (0.94)
Moving average ± 5 years 0.0014 (0.47) 0.0012 (0.39) 0.0013 (0.42) 0.0014 (0.78) 0.0014 (0.78) 0.0014 (0.78)
Moving average ± 10 years 0.0017 (0.56) 0.0015 (0.48) 0.0016 (0.52) 0.0016 (0.89) 0.0016 (0.89) 0.0017 (0.94)
Random walk 70.0005 (0.16)* 0.0004 (0.13)* 0.0006 (0.19)* 70.0004 (0.22)* 0.0004 (0.22)* 0.0005 (0.28)*
Exponential smoothing 0.0016 (0.53) 0.0016 (0.52) 0.0017 (0.55) 0.0017 (0.94) 0.0017 (0.94) 0.0017 (0.94)
Recursive exponential smoothing 0.0008 (0.26) 0.0008 (0.26) 0.0010 (0.32) 0.0017 (0.94) 0.0017 (0.94) 0.0017 (0.94)
EWMA 0.0018 (0.58) 0.0020 (0.65) 0.0019 (0.61) 0.0018 (1.00) 0.0017 (0.94) 0.0017 (0.94)
Recursive EWMA 0.0017 (0.55) 0.0017 (0.55) 0.0018 (0.58) 0.0014 (0.78) 0.0014 (0.78) 0.0015 (0.83)
Simple regression 0.0028 (0.89) 0.0028 (0.90) 0.0028 (0.90) 0.0019 (1.06) 0.0019 (1.06) 0.0019 (1.06)
Recursive simple regression 0.0027 (0.87) 0.0027 (0.87) 0.0027 (0.87) 0.0017 (0.94) 0.0017 (0.94) 0.0017 (0.94)
GARCH 0.0015 (0.47) 0.0015 (0.48) 0.0015 (0.48) 0.0029 (1.64) 0.0029 (1.64) 0.0030 (1.67)
Recursive GARCH 0.0013 (0.42) 0.0013 (0.42) 0.0014 (0.47) 0.0016 (0.89) 0.0016 (0.89) 0.0018 (1.00)
TGARCH 0.0017 (0.56) 0.0017 (0.55) 0.0018 (0.58) 0.0065 (3.72) 0.0065 (3.72) 0.0067 (3.72)
Recursive TGARCH 0.0016 (0.51) 0.0016 (0.51) 0.0018 (0.58) 0.0022 (1.22) 0.0022 (1.22) 0.0024 (1.33)
EGARCH 0.0020 (0.64) 0.0020 (0.65) 0.0021 (0.68) 0.0023 (1.28) 0.0023 (1.28) 0.0024 (1.33)
Recursive EGARCH 0.0017 (0.55) 0.0017 (0.55) 0.0019 (0.61) 0.0019 (1.06) 0.0019 (1.06) 0.0019 (1.06)
C-GARCH 0.0016 (0.51) 0.0016 (0.52) 0.0016 (0.52) 0.0022 (1.22) 0.0022 (1.22) 0.0022 (1.22)
Recursive C-GARCH 0.0015 (0.47) 0.0015 (0.47) 0.0016 (0.51) 0.0018 (1.00) 0.0018 (1.00) 0.0019 (1.06)

Notes: FTA-ALL refers to the Financial Times-Actuaries All Share Index, and FTSE to the Financial Times± Stock Exchange 100 index;
ME is the mean error statistic de® ned in (17). MAE is the mean absolute error statistic de® ned in (18). RMSE is the root mean squared
error statistic de® ned in (19). Relative error statistics obtained by expressing the actual statistic for each model as a ratio to the
corresponding error statistic for the historical mean are provided in parentheses, minimum values for which are indicated by asterisks.
All forecast error statistics relate to the period 1/1/1995± 31/7/1996. For forecasting model descriptions and de® nitions see Section III of
the text, expressions 3± 15.

three-month moving average is marginally superior to the are far superior (using RMSE) to the historical mean and
random walk and other (particularly recursive) smoothing simple regression models which again provide the poorest
models. On the basis of the RMSE criterion there is far less forecasts. Moreover, the random walk model performs poorly
divergence in error statistics for both series for the majority of on the RMSE measure, both relative to the results for lower
models considered, although the 3-month moving average frequencies noted above and to other models for daily data,
model is marginally favoured for both FTA and FTSE data and for (particularly crash-adjusted ) FTSE data especially.
crash-unadjusted, exponential smoothing for the FTA In sum, on the basis of symmetric forecast error statis-
adjusted series, and the GARCH model for the crash- tics, while the random walk model provides the most accu-
adjusted FTSE series. For all daily series, all of these models rate volatility forecasts at the monthly and weekly
 444 D. McMillan et al.
Table 2. Forecast error statistics; weekly frequency

FTA-All FTSE

Model ME MAE RMSE ME MAE RMSE

A. Excluding dummy for 1987 crash

Historical mean 0.00042 (1.00) 0.00050 (1.00) 0.00053 (1.00) 0.00037 (1.00) 0.00040 (1.00) 0.00044 (1.00)
Recursive historical mean 0.00045 (1.07) 0.00049 (0.98) 0.00051 (0.96) 0.00034 (0.92) 0.00038 (0.95) 0.00042 (0.95)
Moving average ± 1.25 years 0.00005 (0.11) 0.00020 (0.40) 0.00025 (0.47)* 0.00010 (0.28)* 0.00020 (0.50) 0.00027 (0.61)*
Moving average ± 2.5 years 0.00005 (0.13) 0.00020 (0.40) 0.00025 (0.47)* 0.00012 (0.33) 0.00021 (0.53) 0.00027 (0.61)*
Random walk 70.00016 (0.37) 0.00016 (0.32)* 0.00029 (0.55) 70.00016 (0.45) 0.00016 (0.40)* 0.00029 (0.66)
Exponential smoothing 0.00015 (0.36) 0.00026 (0.52) 0.00029 (0.55) 0.00043 (1.17) 0.00046 (1.15) 0.00049 (1.11)
Recursive exponential smoothing 0.00002 (0.06)* 0.00019 (0.38) 0.00025 (0.47)* 0.00036 (0.97) 0.00040 (1.00) 0.00044 (1.00)
EWMA 0.00005 (0.13) 0.00025 (0.50) 0.00029 (0.55) 0.00017 (0.47) 0.00024 (0.60) 0.00030 (0.68)
Recursive EWMA 0.00005 (0.13) 0.00020 (0.40) 0.00025 (0.47)* 0.00012 (0.32) 0.00021 (0.53) 0.00027 (0.61)*
Simple regression 0.00055 (1.09) 0.00047 (0.94) 0.00049 (0.93) 0.00036 (0.99) 0.00041 (1.03) 0.00044 (1.00)
Recursive simple regression 0.00042 (1.00) 0.00046 (0.92) 0.00048 (0.91) 0.00034 (0.92) 0.00038 (0.95) 0.00042 (0.95)
GARCH 0.00021 (0.50) 0.00029 (0.58) 0.00033 (0.62) NA NA NA
Downloaded by [The Aga Khan University] at 21:40 23 December 2014

Recursive GARCH 0.00020 (0.48) 0.00028 (0.56) 0.00032 (0.60) NA NA NA

TGARCH 0.00021 (0.49) 0.00029 (0.58) 0.00032 (0.60) NA NA NA
Recursive TGARCH 0.00019 (0.42) 0.00028 (0.56) 0.00031 (0.58) NA NA NA
EGARCH 0.00013 (0.31) 0.00024 (0.48) 0.00029 (0.55) NA NA NA
Recursive EGARCH 0.00012 (0.45) 0.00024 (0.48) 0.00028 (0.53) NA NA NA
C-GARCH 0.00021 (0.50) 0.00029 (0.58) 0.00033 (0.62) NA NA NA
Recursive C-GARCH 0.00020 (0.42) 0.00028 (0.56) 0.00032 (0.60) NA NA NA

B. Including dummy for 1987 crash

Historical mean 0.00040 (1.00) 0.00047 (1.00) 0.00049 (1.00) 0.00027 (1.00) 0.00032 (1.00) 0.00036 (1.00)
Recursive historical mean 0.00042 (1.05) 0.00046 (0.98) 0.00048 (0.98) 0.00026 (0.96) 0.00030 (0.94) 0.00034 (0.94)
Moving average ± 1.25 years 0.00005 (0.12) 0.00020 (0.43) 0.00025 (0.51)* 0.00010 (0.38)* 0.00020 (0.63) 0.00025 (0.69)*
Moving average ± 2.5 years 0.00005 (0.13) 0.00019 (0.40) 0.00025 (0.51)* 0.00013 (0.46) 0.00021 (0.66) 0.00026 (0.72)
Random walk 70.00016 (0.38) 0.00016 (0.34)* 0.00029 (0.59) 70.00016 (0.57) 0.00016 (0.50)* 0.00027 (0.75)
Exponential smoothing 0.00017 (0.42) 0.00026 (0.55) 0.00030 (0.61) 0.00028 (1.03) 0.00033 (1.03) 0.00036 (1.00)
Recursive exponential smoothing 0.00001 (0.03)* 0.00018 (0.38) 0.00025 (0.51)* 0.00012 (0.44) 0.00022 (0.69) 0.00027 (0.75)
EWMA 0.00014 (0.35) 0.00025 (0.53) 0.00028 (0.57) 0.00018 (0.66) 0.00025 (0.78) 0.00029 (0.81)
Recursive EWMA 0.00006 (0.15) 0.00020 (0.43) 0.00025 (0.51)* 0.00013 (0.48) 0.00021 (0.66) 0.00026 (0.72)
Simple regression 0.00037 (0.92) 0.00042 (0.89) 0.00045 (0.92) 0.00027 (0.97) 0.00031 (0.97) 0.00035 (0.97)
Recursive simple regression 0.00037 (0.93) 0.00041 (0.87) 0.00044 (0.90) 0.00025 (0.93) 0.00030 (0.94) 0.00034 (0.94)
GARCH 0.00014 (0.34) 0.00025 (0.53) 0.00029 (0.59) 0.00016 (0.59) 0.00024 (0.75) 0.00028 (0.78)
Recursive GARCH 0.00013 (0.33) 0.00024 (0.51) 0.00028 (0.57) 0.00013 (0.48) 0.00022 (0.69) 0.00026 (0.72)
TGARCH 0.00013 (0.31) 0.00024 (0.51) 0.00028 (0.57) 0.00018 (0.67) 0.00025 (0.78) 0.00029 (0.81)
Recursive TGARCH 0.00011 (0.28) 0.00023 (0.49) 0.00027 (0.55) 0.00015 (0.56) 0.00023 (0.72) 0.00027 (0.75)
EGARCH 0.00009 (0.22) 0.00022 (0.47) 0.00027 (0.55) 0.00017 (0.62) 0.00024 (0.75) 0.00029 (0.81)
Recursive EGARCH 0.00008 (0.20) 0.00021 (0.45) 0.00027 (0.55) 0.00014 (0.52) 0.00022 (0.69) 0.00027 (0.75)
C-GARCH 0.00014 (0.35) 0.00025 (0.53) 0.00029 (0.59) 0.00015 (0.53) 0.00023 (0.72) 0.00028 (0.78)
Recursive C-GARCH 0.00012 (0.30) 0.00024 (0.51) 0.00028 (0.57) 0.00013 (0.48) 0.00022 (0.69) 0.00027 (0.75)

Note: As Table 1.
frequencies (especially monthly, where it dominates all and historical mean models, ranks poorly compared to the
others by a sizeable margin), it performs poorly in forecast- above models.
ing the volatility of daily returns. The exponential smooth-
ing and moving average models provide some of the most
accurate weekly and daily volatility forecasts, but poorer Asymmetric forecast error results
monthly volatility forecasts, the moving average models
Finally, Tables 4 and 5 report MME(U) and MME(O)
providing a consistently good relative forecasting perform-
statistics for all series at the weekly and daily frequencies
ance. Similarly, the GARCH genre of models provide a respectively. No such statistics are reported for monthly
consistently fair relative forecasting performance, series, where all models other than the random walk con-
GARCH and EGARCH models outperforming the sistently overpredict volatility throughout the out-of-
TGARCH and CGARCH models. The performance of sample data, (the MME(U) statistic in particular collapsing
the remaining models, and especially the simple regression to the MAE statistic reported above in such circum-
stances). At the daily frequency, for both the FTA and
 Forecasting UK stock market volatility 445
Table 3. Forecast error statistics; daily frequency

FTA-All FTSE

Model ME MAE RMSE ME MAE RMSE

A. Excluding dummy for 1987 crash

Historical mean 7.34e-05 (1.00) 7.87e-05 (1.00) 8.32e-05 (1.00) 5.41e-05 (1.00) 6.74e-05 (1.00) 7.48e-05 (1.00)
Recursive historical mean 7.13e-05 (0.97) 7.67e-05 (0.97) 8.14e-05 (0.98) 5.06e-05 (0.94) 6.45e-05 (0.96) 7.23e-05 (0.97)
Moving average ± 3 months 1.58e-07 (0.02) 2.45e-05 (0.31) 3.93e-05 (0.47)* 7.96e-07 (0.01) 3.45e-05 (0.51) 5.19e-05 (0.69)*
Moving average ± 6 months 2.88e-06 (0.04) 2.56e-05 (0.33) 3.98e-05 (0.48) 4.29e-06 (0.08) 3.57e-05 (0.53) 5.24e-05 (0.70)
Random walk 72.54e-05 (0.35) 2.56e-05 (0.33) 4.68e-05 (0.56) 73.57e-05 (0.66) 3.59e-05 (0.53) 6.30e-05 (0.84)
Exponential smoothing 73.56e-05 (0.05) 2.34e-05 (0.30)* 3.94e-05 (0.47) 71.98e-05 (0.36) 3.09e-05 (0.46)* 5.54e-05 (0.74)
Recursive exponential smoothing 71.17e-07 (0.002)* 2.65e-05 (0.34) 4.17e-05 (0.50) 71.42e-07 (0.003)* 4.31e-05 (0.64) 6.74e-05 (0.90)
EWMA 2.30e-05 (0.31) 3.70e-05 (0.47) 4.95e-05 (0.60) 3.02e-05 (0.56) 4.96e-05 (0.74) 5.99e-05 (0.80)
Recursive EWMA 2.94e-06 (0.04) 2.58e-05 (0.33) 4.01e-05 (0.48) 4.36e-06 (0.08) 3.60e-05 (0.53) 5.27e-05 (0.70)
Simple regression 3.91e-05 (0.53) 4.96e-05 (0.63) 5.81e-05 (0.70) 1.97e-05 (0.36) 4.65e-05 (0.69) 6.47e-05 (0.87)
Recursive simple regression 3.79e-05 (0.52) 4.86e-05 (0.62) 5.74e-05 (0.69) 1.85e-05 (0.34) 4.58e-05 (0.68) 6.43e-05 (0.86)
GARCH 1.36e-05 (0.19) 3.13e-05 (0.40) 4.20e-05 (0.51) 2.08e-05 (0.38) 4.42e-05 (0.66) 5.63e-05 (0.75)
Downloaded by [The Aga Khan University] at 21:40 23 December 2014

Recursive GARCH 1.31e-05 (0.18) 3.10e-05 (0.39) 4.19e-05 (0.50) 1.90e-05 (0.35) 4.31e-05 (0.64) 5.56e-05 (0.74)
TGARCH 1.46e-05 (0.20) 3.19e-05 (0.41) 4.23e-05 (0.51) 2.07e-05 (0.38) 4.39e-05 (0.65) 5.59e-05 (0.75)
Recursive TGARCH 1.36e-05 (0.19) 3.12e-05 (0.40) 4.20e-05 (0.50) 1.89e-05 (0.35) 4.29e-05 (0.64) 5.53e-05 (0.74)
EGARCH 8.40e-06 (0.11) 2.86e-05 (0.36) 4.07e-05 (0.49) 1.97e-05 (0.36) 4.35e-05 (0.65) 5.59e-05 (0.75)
Recursive EGARCH 9.03e-06 (0.12) 2.89e-05 (0.36) 4.07e-05 (0.49) 1.76e-05 (0.33) 4.26e-05 (0.63) 5.53e-05 (0.74)
C-GARCH 1.15e-05 (0.16) 3.02e-05 (0.38) 4.15e-05 (0.50) 2.02e-05 (0.37) 4.39e-05 (0.65) 5.61e-05 (0.75)
Recursive C-GARCH 1.02e-05 (0.14) 2.93e-05 (0.37) 4.12e-05 (0.50) 1.67e-05 (0.31) 4.19e-05 (0.62) 5.49e-05 (0.73)

B. Including dummy for 1987 crash

Historical mean 7.05e-05 (1.00) 7.60e-05 (1.00) 8.07e-05 (1.00) 4.46e-05 (1.00) 5.98e-05 (1.00) 6.84e-05 (1.00)
Recursive historical mean 6.84e-05 (0.97) 7.41e-05 (0.98) 7.89e-05 (0.98) 4.17e-05 (0.93) 5.76e-05 (0.96) 6.69e-05 (0.97)
Moving average ± 3 months 1.88e-07 (0.003)* 2.45e-05 (0.32)* 3.94e-05 (0.49) 9.14e-07 (0.02) 3.45e-05 (0.58)* 5.25e-05 (0.77)
Moving average ± 6 months 2.88e-06 (0.04) 2.55e-05 (0.34) 3.98e-05 (0.49) 4.40e-06 (0.10) 3.56e-05 (0.60) 5.24e-05 (0.77)
Random walk 72.53e-05 (0.36) 2.55e-05 (0.34) 4.68e-05 (0.58) 73.52e-05 (0.79) 3.54e-05 (0.59) 6.32e-05 (0.92)
Exponential smoothing 1.80e-06 (0.03) 2.51e-05 (0.33) 3.93e-05 (0.49)* 3.33e-06 (0.07) 3.54e-05 (0.59) 5.24e-05 (0.77)
Recursive exponential smoothing 72.53e-07 (0.004) 2.57e-05 (0.34) 4.06e-05 (0.50) 71.57e-07 (0.004)* 3.58e-05 (0.60) 5.41e-05 (0.79)
EWMA 2.29e-05 (0.32) 3.69e-05 (0.49) 4.55e-05 (0.56) 3.00e-05 (0.67) 4.92e-05 (0.82) 5.98e-05 (0.87)
Recursive EWMA 2.94e-06 (0.04) 2.57e-05 (0.34) 4.01e-05 (0.50) 4.47e-06 (0.10) 3.60e-05 (0.60) 5.33e-05 (0.78)
Simple regression 4.61e-05 (0.65) 5.50e-05 (0.72) 6.20e-05 (0.77) 3.22e-05 (0.72) 5.16e-05 (0.86) 6.34e-05 (0.93)
Recursive simple regression 4.46e-05 (0.63) 5.38e-05 (0.71) 6.09e-05 (0.75) 3.02e-05 (0.68) 5.01e-05 (0.84) 6.24e-05 (0.91)
GARCH 1.27e-05 (0.18) 3.08e-05 (0.41) 4.19e-05 (0.52) 1.59e-05 (0.36) 4.19e-05 (0.70) 5.23e-05 (0.77)*
recursive GARCH 1.19e-05 (0.17) 3.04e-05 (0.40) 4.17e-05 (0.52) 1.58e-05 (0.35) 4.16e-05 (0.70) 5.52e-05 (0.80)
TGARCH 1.21e-05 (0.17) 3.04e-05 (0.40) 4.15e-05 (0.51) 1.74e-05 (0.39) 4.22e-05 (0.71) 5.54e-05 (0.81)
Recursive TGARCH 1.14e-05 (0.16) 3.00e-05 (0.39) 4.14e-05 (0.51) 1.56e-05 (0.35) 4.14e-05 (0.69) 5.50e-05 (0.80)
EGARCH 8.81e-06 (0.12) 2.88e-05 (0.38) 4.09e-05 (0.51) 1.59e-05 (0.36) 4.19e-05 (0.70) 5.54e-05 (0.81)
Recursive EGARCH 8.34e-06 (0.12) 2.86e-05 (0.38) 4.07e-05 (0.50) 1.43e-05 (0.32) 4.11e-05 (0.69) 5.50e-05 (0.80)
C-GARCH 9.90e-06 (0.14) 2.94e-05 (0.39) 4.11e-05 (0.51) 1.73e-05 (0.39) 4.21e-05 (0.70) 5.50e-05 (0.80)
Recursive C-GARCH 8.70e-06 (0.00) 2.88e-05 (0.38) 4.09e-05 (0.51) 1.53e-05 (0.34) 4.13e-05 (0.69) 5.50e-05 (0.80)

Note: As Table 1.

FTSE series adjusted or unadjusted for the 1987 crash, the V I . C ON C L U S I ON

historical (or recursive historical) mean is the best volatility
forecasting model if underpredictions are penalized more This paper provides a comparative evaluation of the ability
heavily, while the random walk model is massively pre- of a variety of statistical and econometric models to fore-
ferred if overpredictions are penalized more heavily. cast the volatility of the UK FTA All Share and FTSE100
These conclusions also hold for the FTA at the weekly stock indices, motivated by recognition of the practical
frequency, although the simple regression model is also a need for accurate volatility forecasts in areas such as option
preferred model on the MME(U) criterion. For the weekly pricing and the limited empirical evidence available to date
FTSE series, the random walk continues to be heavily for the UK. A total of ten volatility forecasting models are
preferred on the MME(O) criterion penalizing overpredic- considered, including the historical mean, moving average,
tions more heavily, while the exponential smoothing model random walk, exponential smoothing, exponentially
is preferred on the MME(U) criterion. weighted moving average, simple regression, GARCH,
 Downloaded by [The Aga Khan University] at 21:40 23 December 2014

446

Table 4. Mean mixed forecast error statistics ± weekly frequency

FTA-All FTSE

Excluding 1987 crash dummy Including 1987 crash dummy Excluding 1987 crash dummy Including 1987 crash dummy

Model MME(U) MME(O) MME(U) MME(O) MME(U) MME(O) MME(U) MME(O)

Historical mean 0.0014 (1.00)* 0.0211 (1.00) 0.0014 (1.00)* 0.0200 (1.00) 0.0012 (1.00) 0.0188 (1.00) 0.0014 (1.00) 0.0161 (1.00)
Recursive historical mean 0.0014 (1.00)* 0.0208 (0.99) 0.0014 (1.00)* 0.0200 (1.00) 0.0012 (1.00) 0.0181 (0.96) 0.0015 (1.07) 0.0154 (0.96)
Moving average ± 1.25 years 0.0042 (3.00) 0.0094 (0.45) 0.0041 (2.93) 0.0093 (0.47) 0.0030 (2.50) 0.0105 (0.56) 0.0028 (2.00) 0.0105 (0.65)
Moving average ± 2.5 years 0.0040 (2.86) 0.0095 (0.45) 0.0039 (2.79) 0.0094 (0.47) 0.0028 (2.33) 0.0113 (0.60) 0.0026 (1.86) 0.0114 (0.71)
Random walk 0.0096 (6.86) 0.0002 (0.01)* 0.0094 (6.71) 0.0002 (0.01)* 0.0105 (8.75) 0.0002 (0.01)* 0.0102 (7.29) 0.0002 (0.01)*
Exponential smoothing 0.0030 (2.14) 0.0126 (0.60) 0.0027 (1.93) 0.0129 (0.65) 0.0010 (0.83)* 0.0205 (1.09) 0.0013 (0.93)* 0.0163 (1.01)
Recursive exponential 0.0045 (3.21) 0.0086 (0.41) 0.0046 (3.29) 0.0082 (0.41) 0.0012 (1.00) 0.0186 (0.99) 0.0028 (2.00) 0.0111 (0.69)
smoothing
EWMA 0.0040 (2.86) 0.0095 (0.45) 0.0031 (2.21) 0.0122 (0.61) 0.0023 (1.92) 0.0129 (0.69) 0.0021 (1.50) 0.0131 (0.81)
Recursive EWMA 0.0040 (2.86) 0.0095 (0.45) 0.0040 (2.86) 0.0096 (0.48) 0.0028 (2.33) 0.0113 (0.60) 0.0026 (1.86) 0.0114 (0.71)
Simple regression 0.0014 (1.00)* 0.0202 (0.96) 0.0014 (1.00)* 0.0187 (0.94) 0.0012 (1.00) 0.0188 (1.00) 0.0014 (1.00) 0.0158 (0.98)
Recursive simple regression 0.0014 (1.00)* 0.0199 (0.94) 0.0015 (1.07) 0.0186 (0.93) 0.0012 (1.00) 0.0181 (0.96) 0.0016 (1.14) 0.0151 (0.94)
GARCH 0.0024 (1.71) 0.0142 (0.67) 0.0031 (2.21) 0.0121 (0.61) NA NA 0.0023 (1.64) 0.0126 (0.78)
Recursive GARCH 0.0025 (1.86) 0.0138 (0.65) 0.0033 (2.36) 0.0119 (0.60) NA NA 0.0026 (1.86) 0.0115 (0.71)
TGARCH 0.0024 (1.71) 0.0141 (0.67) 0.0031 (2.21) 0.0117 (0.59) NA NA 0.0022 (1.57) 0.0131 (0.81)
Recursive TGARCH 0.0026 (1.86) 0.0136 (0.64) 0.0033 (2.36) 0.0115 (0.58) NA NA 0.0024 (1.71) 0.0120 (0.75)
EGARCH 0.0032 (2.29) 0.0118 (0.56) 0.0036 (2.57) 0.0104 (0.52) NA NA 0.0023 (1.64) 0.0128 (0.80)
Recursive EGARCH 0.0033 (2.36) 0.0115 (0.55) 0.0037 (2.64) 0.0103 (0.52) NA NA 0.0026 (1.86) 0.0118 (0.73)
C-GARCH 0.0024 (1.71) 0.0142 (0.67) 0.0031 (2.21) 0.0121 (0.61) NA NA 0.0025 (1.79) 0.0120 (0.75)
Recursive C-GARCH 0.0025 (1.79) 0.0138 (0.65) 0.0033 (2.36) 0.0118 (0.59) NA NA 0.0027 (1.93) 0.0114 (0.71)

Notes: As Table 1 additionally, MME(U) is the asymmetric mean error statistic de® ned in Equation 20 which penalizes underpredictions more heavily than overpredictions ;
and MME(O) is the asymmetric mean error statistic de® ned in Equation 21 which penalizes overpredictions more heavily.
D. McMillan et al.
 Downloaded by [The Aga Khan University] at 21:40 23 December 2014

Table 5. Mean mixed forecast error statistics ± daily frequency

FTA-All FTSE

Excluding 1987 crash dummy Including 1987 crash dummy Excluding 1987 crash dummy Including 1987 crash dummy
Forecasting UK stock market volatility

Model MME(U) MME(O) MME(U) MME(O) MME(U) MME(O) MME(U) MME(O)

Historical mean 0.0004 (1.00)* 0.0084 (1.00) 0.0004 (1.00)* 0.0082 (1.00) 0.0008 (1.00)* 0.0072 (1.00) 0.0009 (1.00)* 0.0066 (1.00)
Recursive historical mean 0.0004 (1.00)* 0.0083 (0.99) 0.0004 (1.00)* 0.0081 (0.99) 0.0008 (1.00)* 0.0070 (0.97) 0.0009 (1.00)* 0.0064 (0.97)
Moving average ± 3 months 0.0017 (4.25) 0.0028 (0.33) 0.0017 (4.25) 0.0028 (0.34) 0.0021 (2.63) 0.0033 (0.46) 0.0021 (2.33) 0.0033 (0.50)
Moving average ± 6 months 0.0016 (4.00) 0.0030 (0.36) 0.0016 (4.00) 0.0030 (0.37) 0.0019 (2.38) 0.0036 (0.50) 0.0019 (2.11) 0.0036 (0.55)
Eandom walk 0.0040 (10.0) 0.00002 (0.002)* 0.0040 (10.0) 0.00002 (0.002)* 0.0048 (6.00) 0.00004 (0.005)* 0.0047 (5.22) 0.0001 (0.02)*
Exponential smoothing 0.0019 (4.75) 0.0025 (0.30) 0.0016 (4.00) 0.0030 (0.37) 0.0032 (4.00) 0.0016 (0.22) 0.0019 (2.11) 0.0036 (0.55)
Recursive exponential 0.0019 (4.75) 0.0028 (0.33) 0.0018 (4.50) 0.0028 (0.34) 0.0027 (3.38) 0.0030 (0.42) 0.0022 (2.44) 0.0033 (0.50)
smoothing
EWMA 0.0009 (2.25) 0.0049 (0.58) 0.0009 (2.25) 0.0049 (0.60) 0.0011 (1.38) 0.0056 (0.78) 0.0011 (1.22) 0.0056 (0.85)
Recursie EWMA 0.0016 (4.00) 0.0030 (0.36) 0.0016 (4.00) 0.0030 (0.37) 0.0019 (2.38) 0.0036 (0.50) 0.0019 (2.11) 0.0036 (0.55)
Simple regression 0.0007 (1.75) 0.0061 (0.73) 0.0006 (1.50) 0.0066 (0.80) 0.0016 (2.00) 0.0046 (0.64) 0.0011 (1.22) 0.0057 (0.86)
Recursive simple regression 0.0007 (1.75) 0.0060 (0.71) 0.0006 (1.50) 0.0065 (0.79) 0.0017 (2.13) 0.0045 (0.63) 0.0012 (1.33) 0.0055 (0.83)
GARCH 0.0012 (3.00) 0.0041 (0.49) 0.0012 (3.00) 0.0040 (0.49) 0.0014 (1.75) 0.0049 (0.68) 0.0016 (1.78) 0.0045 (0.68)
Recursive GARCH 0.0012 (3.00) 0.0040 (0.48) 0.0013 (3.25) 0.0039 (0.48) 0.0015 (1.88) 0.0047 (0.65) 0.0016 (1.78) 0.0045 (0.68)
TGARCH 0.0012 (3.00) 0.0042 (0.50) 0.0013 (3.25) 0.0039 (0.48) 0.0014 (1.75) 0.0049 (0.68) 0.0015 (1.67) 0.0046 (0.70)
Recursive TGARCH 0.0012 (3.00) 0.0041 (0.49) 0.0013 (3.25) 0.0039 (0.48) 0.0015 (1.88) 0.0047 (0.65) 0.0016 (1.78) 0.0045 (0.68)
EGARCH 0.0014 (3.50) 0.0036 (0.43) 0.0014 (3.50) 0.0036 (0.44) 0.0015 (1.88) 0.0048 (0.67) 0.0016 (1.78) 0.0045 (0.68)
Recursive EGARCH 0.0014 (3.50) 0.0036 (0.43) 0.0014 (3.50) 0.0036 (0.44) 0.0015 (1.88) 0.0047 (0.65) 0.0016 (1.78) 0.0044 (0.67)
C-GARCH 0.0013 (3.25) 0.0039 (0.46) 0.0014 (3.50) 0.0037 (0.45) 0.0014 (1.75) 0.0048 (0.67) 0.0015 (1.67) 0.0046 (0.70)
Recursive C-GARCH 0.0013 (3.25) 0.0037 (0.44) 0.0014 (3.50) 0.0036 (0.44) 0.0015 (1.88) 0.0046 (0.64) 0.0016 (1.78) 0.0045 (0.68)

Note: as Table 4.
447
 448 D. McMillan et al.
TGARCH, EGARCH, and component-GARCH models, Bollerslev, T., Chou, R. Y. and Kroner, K. F. (1992) ARCH
and, additionally, recursive variants of these models where modelling in ® nance : a review of the theory and empirical
evidence, Journal of Econometrics, 52, 61± 90.
appropriate. Forecast evaluations are performed for Brailsford, T. J. and Fa€ , R. W. (1996 ) An evaluation of volatility
monthly, weekly and daily data frequencies, both with forecasting techniques, Journal of Banking and Finance, 20,
and without adjustment for the 1987 crash, and with 419 ± 38.
respect to both symmetric and asymmetric loss functions. de Jong, F., Kemna, A. and Kloeck, T. (1992) A contribution to
In summary of our results, when asymmetric loss is con- event study methodology with an application to the Dutch
sidered, the ranking of forecasting methods is dependent stock market, Journal of Banking and Finance, 16, 11± 36.
Dimson, E. and Marsh, P. (1990) Volatility forecasting without
on the series, frequency and direction of that asymmetry, in data-snooping, Journal of Banking and Finance, 14, 399± 421.
keeping with the previous results of Brailsford and Fa€ Edwards, F. R. (1988a) Does Futures trading increase stock mar-
(1996). If overpredictions are penalized more heavily than ket volatility? Financial Analysts Journal, 44, 63± 9.
under predictions, the random walk model is favoured. Edwards, F. R. (1988b) Futures trading and cash market volati-
However, if underpredictions are penalized more heavily lity: stock index and interest rate futures, Journal of Futures
Markets, 8, 421± 39.
than overpredictions, then the historical mean is favoured
Engle, R. F. (1982) Autoregressive conditional heteroscedasticity
for the forecasting of daily FTA and FTSE volatility, while with estimates of the variance of UK in¯ ation, Econometrica,
Downloaded by [The Aga Khan University] at 21:40 23 December 2014

the historical mean and simple regression are jointly 50, 987 ± 1007.
favoured for weekly FTA volatility, and exponential Engle, R. F. and Lee, G. G. J. (1993) A permanent and transitory
smoothing is favoured for forecasting weekly FTSE vola- component model of stock return volatility, UCSD,
tility. For the symmetric loss case, the random walk model Department of Economics, Discussion Paper No: 92-44R .
Engle, R. F. and Ng, V. K. (1993) Measuring and testing the
is found to provide vastly superior monthly volatility fore- impact of news on volatility, Journal of Finance, 48, 1749 ± 78.
casts, while the random walk, moving average, and recur- Fama, E. F. (1965) The behaviour of stock market prices, Journal
sive smoothing models provide moderately superior weekly of Business, 38, 34± 105.
volatility forecasts, and GARCH, moving average, and Franses, P. H. and Dijk, D. van (1996) Forecasting stock market
exponential smoothing models provide marginally superior volatility using (non-linear) GARCH models, Journal of
daily volatility forecasts. Our ® ndings also lend some sup- Forecasting, 15, 229 ± 35.
Gemmill, G. (1993) Options Pricing, McGraw-Hill, Maidenhead.
port to the results of Franses and van Dijk (1996) concern- Glosten, L. R., Jagannathan, R. and Runkle, D. E. (1993) On the
ing the dominance of the random walk model when the relation between the expected value and volatility of the nom-
crash of 1987 is included in the estimation sample, and the inal excess return on stocks, Journal of Finance, 48, 1779± 801.
improvement in GARCH model forecasts obtained follow- Kearns, P. and Pagan, A. R. (1993) Australian stock market
ing its exclusion. Indeed, if attention is restricted to sym- volatility: 1857± 1987, T he Economic Record, 69, 163 ± 78.
Mandelbrot, B. (1963) The variation of certain speculative prices,
metric loss and the proposal of one forecasting method for
Journal of Business, 36, 394 ± 419.
all frequencies, the most consistent forecasting performance Nelson, D. B. (1991) Conditional heteroscedasticity in asset
is provided by moving average and GARCH models. More returns : A new approach, Econometrica , 59, 347 ± 70.
generally, our results suggest that previous results reporting Nelson, D. B. and Cao, C. Q. (1992) Inequality constraints in the
that the class of GARCH models provide relatively poor univariate GARCH model, Journal of Business and Economic
volatility forecasts, so limiting their practical usefulness, Statistics, 10, 229± 35.
Pagan, A. R. and Schwert, G. W. (1990) Alternative models for
may be data speci® c, holding more robustly at lower fre- conditional stock market volatility, Journal of Econometrics,
quencies but failing to hold at higher frequencies, as here 45, 267 ± 90.
for the crash-adjusted daily FTSE100 index in particular. Plummer, T. (1989) Forecasting Financial Markets : The T ruth
Behind T echnical Analysis, Kogan Page, London.
Robinson, G. (1994) The e€ ects of futures trading on cash market
R E F E R E N C ES volatility; evidence from the London stock exchange, Review
Akgiray, V. (1989) Conditional heteroscedasticity in time series of of Futures Markets, 13, 429 ± 52.
stock returns, Journal of Business, 62, 55± 80. Schwert, G. W. (1990) Stock market volatility, Financial Analysts
Antoniou, A. and Holmes, P. (1995) Futures trading, information Journal, 46, 23± 34.
and spot price volatility: Evidence for the FTSE100 stock Stewart, T. H. (1966) How Charts Can Make you Money: Technical
index futures contract using GARCH, Journal of Banking Analysis For Investors, Woodhead-Faulkner, Cambridge.
and Finance, 19, 117± 29. Theobald, M. and Price, V. (1984) Seasonality estimation in thin
Beveridge, S. and Nelson, C. R. (1981) A new approach to decom- markets, Journal of Finance, 39, 377 ± 92.
position of economic time series into permanent and transi- Tse, Y. K. (1991) Stock return volatility in the Tokyo stock
tory components with particular attention to measurement of exchange, Japan and the W orld Econom y, 3, 285 ± 98.
the business cycle, Journal of Monetary Economics, 7, 151 ± 74. Tse, Y. K. and Tung, S. H. (1992) Forecasting volatility in the
Black, F. (1976) Studies in stock price volatility changes, Singapore stock market, Asia Paci® c Journal of Management,
Proceedings of the 1976 Business and Economics Statistics 9, 1± 13.
Section, American Statistical Association, pp. 177± 81. Vasilellis, G. A. and Meade, N. (1996 ) Forecasting volatility for
Bollerslev, T. (1986 ) Generalized autoregressive conditional het- portfolio selection, Journal of Business Finance and
eroscedasticity, Journal of Econometrics, 31, 307± 27. Accounting, 23, 125 ± 43.