volatility forecasting for risk management
TRANSCRIPT
Journal of ForecastingJ. Forecast. 22, 1–22 (2003)Published online 9 October 2002 in Wiley InterScience (www.interscience.wiley.com). DOI: 10.1002/for.841
Volatility Forecasting for RiskManagement
CHRIS BROOKS1* AND GITA PERSAND2
1 ISMA Centre, University of Reading, UK2 Department of Economics, University of Bristol, UK
ABSTRACTRecent research has suggested that forecast evaluation on the basis of stan-dard statistical loss functions could prefer models which are sub-optimalwhen used in a practical setting. This paper explores a number of statis-tical models for predicting the daily volatility of several key UK financialtime series. The out-of-sample forecasting performance of various linear andGARCH-type models of volatility are compared with forecasts derived froma multivariate approach. The forecasts are evaluated using traditional met-rics, such as mean squared error, and also by how adequately they performin a modern risk management setting. We find that the relative accuraciesof the various methods are highly sensitive to the measure used to eval-uate them. Such results have implications for any econometric time seriesforecasts which are subsequently employed in financial decision making.Copyright 2003 John Wiley & Sons, Ltd.
KEY WORDS internal risk management models; asset return volatility; Valueat Risk models; forecasting; univariate and multivariateGARCH models
INTRODUCTION
Modelling and forecasting stock market volatility has been the subject of a great deal of debate overthe past fifteen years or so. Volatility, usually measured by the standard deviation of portfolio returns,is uniquely important in financial markets, for it is often taken to represent the portfolio’s risk.Consequently, the literature on forecasting volatility is sizeable and still growing. Akgiray (1989),for example, finds the GARCH model superior to ARCH, exponentially weighted moving average,and historical mean models for forecasting monthly US stock index volatility. A similar resultconcerning the apparent superiority of GARCH is observed by West and Cho (1995) using one-step-ahead forecasts of dollar exchange rate volatility, evaluated using root-mean squared predictionerrors. However, for longer horizons, the model behaves no better than their alternatives.1 Also using
* Correspondence to: Chris Brooks, ISMA Centre, PO Box 242, University of Reading, Whiteknights, Reading RG6 6BA, UK. E-mail: [email protected] The alternative models are the long-term mean, IGARCH, autoregressive models, and a non-parametric model based on the Gaussiankernel.
Copyright 2003 John Wiley & Sons, Ltd.
2 C. Brooks and G. Persand
the same models and data, West et al. (1993) use asymmetric, utility-based criteria for evaluatingthe conditional variance forecasts, finding that GARCH models tend to yield the highest utilities.Pagan and Schwert (1990) compare GARCH, EGARCH, Markov switching regime and three non-parametric models for forecasting monthly US stock return volatilities. The EGARCH followedby the GARCH models perform moderately; the remaining models produce very poor predictions.Franses and van Dijk (1996) compare three members of the GARCH family (standard GARCH,QGARCH and the GJR model) for forecasting the weekly volatility of various European stockmarket indices. They find that the non-linear GARCH models were unable to beat the standardGARCH model. Brailsford and Faff (1996) find GJR and GARCH models slightly superior tovarious simpler models2 for predicting Australian monthly stock index volatility. The conclusionarising from this growing body of research is that forecasting volatility is a ‘notoriously difficulttask’ (Brailsford and Faff, 1996, p. 419), although it appears that conditional heteroscedasticitymodels are among the best that are currently available. In particular, more complex non-linear andnon-parametric models are inferior in prediction to simpler models, a result echoed in an earlierpaper by Dimson and Marsh (1990) in the context of relatively complex versus parsimonious linearmodels. Finally Brooks (1998) uses a measure of market volume in volatility forecasting models,but observes no increase in forecasting power.
Recent papers have also sought to compare the predictive ability of volatility forecasts derivedfrom the market prices of traded options, with those generated using econometric models (see, forexample, Heynen and Kat, 1994 or Day and Lewis, 1992). The general consensus appears to be thatimplied volatility forecasts are more accurate than those derived using pure time series analysis,but also that the latter still contain additional information not embedded in the implied values.
Also over the past decade, there has been rapid development of techniques for measuring andmanaging financial risk, partially motivated by a spate of recent financial disasters involving deriva-tive securities. One of the most popular approaches to risk measurement is by calculating what isknown as an institution’s ‘Value at Risk’ (VaR). Broadly speaking, Value at Risk is an estimationof likely losses which could arise from changes in market prices. More precisely, it is defined asthe money-loss in a portfolio that is expected to occur over a pre-determined horizon and with apre-determined degree of confidence. The roots of VaR’s popularity stem from the simplicity of itscalculation, its ease of interpretation, and from the fact that VaR can be suitably aggregated acrossan entire firm to produce a single number which broadly encompasses the risk of the positions of thefirm as a whole. Jorion (1996) or Dowd (1998) provide thorough introductions to VaR, and Brooksand Persand (2000a,b) present recent discussions of VaR model estimation issues. The value at riskestimate is also often known as the position risk requirement or minimum capital risk requirement(MCRR); we shall use the three terms interchangeably in the exposition below.
Although the academic literature has thus far failed to keep pace with this expansion, evi-denced by the relatively few academic studies that address this topic, one exception is the studyby Jackson et al. (1998), which assesses the empirical performance of various models for VaRusing historical returns from the actual portfolio of a large investment bank. They find that non-parametric, simulation-based techniques yield more accurate measures of the tail probabilities thanparametric models. Alexander and Leigh (1997) offer an analysis of the relative performance ofequally weighted, exponentially weighted moving average (EWMA), and GARCH model forecastsof volatility, evaluated using traditional statistical and operational adequacy criteria. The GARCH
2 The other models employed are the random walk, the historical mean, a short- and a long-term moving average, exponential smoothing,an exponentially weighted moving average model, and a linear regression.
Copyright 2003 John Wiley & Sons, Ltd. J. Forecast. 22, 1–22 (2003)
Volatility Forecasting for Risk Management 3
model is found to be preferable to EWMA in terms of minimizing the number of exceedencesin a backtest, although the simple unweighted average is superior to both. Brooks et al. (2002)investigate the effectiveness of various hedging models when assessed according to their ability tominimize VaR, finding that there is a large role for time-varying volatilities and correlations, but avery minor role for asymmetries.
This paper seeks to combine and advance the two literatures in volatility forecasting and financialrisk management in a number of ways. First, the volatility forecasting debate is re-opened, and theforecasts from the various models evaluated on the basis of how well they perform in a modernrisk management setting, as well as by traditional statistical loss functions. This is important forDacco and Satchell (1999) demonstrate that the evaluation of forecasts from non-linear modelsusing statistical measures can be misleading, and they propose the use of alternative economicloss functions. Here, the relative performances of the forecasting models are evaluated using bothstatistical and economic loss functions, so that a comparison can be drawn between the two. Second,we also directly compare the forecasting performance of univariate and multivariate forecastingmodels for financial asset return volatility. Multivariate GARCH models permit the estimationof the conditional covariances between assets’ returns, and explicit modelling of this interactionmay improve the accuracy of forecasts of volatility for a portfolio comprising these components.Finally, we evaluate forecasts over the 1- 5-, 10- and 20-day horizons. Although many volatilityforecasting papers compare accuracies at daily horizons, it is often the case that financial marketpractitioners require predictions of much lower frequency. For example, the Basle Committee onBanking Supervision rules for the use of VaR models (see, for example, Basle Committee onBanking Supervision, 1998) require the use of a 10-day holding period, which allows reasonabletime for investors to unwind a position, and fund managers typically re-balance their portfolios ona monthly (20 trading days) basis.
The remainder of the paper is organized as follows. The next section presents the data employed inthe study, while the forecasting models are described briefly in the third section. Forecast evaluationmethods are outlined and discussed in the fourth section with results given in the fifth section. Thefinal section summarises the paper, and offers some concluding remarks.
THE DATA
In this study we calculate the VaRs for three different assets—the FTSE All Share Total ReturnIndex, the FTA British Government Bond (over 15 years) Index and the Reuters CommoditiesPrice Index, as well as an equally weighted portfolio containing these three assets.3, 4 The datawere collected from Datastream International, and spans the period 1 January 1980 to 25 March1999. Observations corresponding to UK public holidays were deleted from the data set to avoidthe incorporation of spurious zero returns, leaving 4865 observations, or trading days in the sample.In the empirical work below, we use the daily log return of the original indices. Summary statisticsfor the data are given in Table I. It is evident that the FTSE returns series is the most volatile,
3 Our analysis assumes that we are long all the three assets—both individually and in the portfolio. A similar analysis could be undertakenfor short or netted positions, but we would not expect our conclusions to be markedly altered.4 This portfolio is deliberately highly simplistic relative to a genuine bank’s book, as well as being entirely linear in nature. The use ofa simple portfolio enables us to more easily unravel the various estimation issues and broad aspects of the methodologies. Additionally,the three series that we consider are all fundamental or ‘benchmark’ factor series, to which other series are mapped under the JP Morganapproach.
Copyright 2003 John Wiley & Sons, Ltd. J. Forecast. 22, 1–22 (2003)
4 C. Brooks and G. Persand
Table I. Summary statistics
Long Govt Bond FTSE All Share Reuters Commodities Portfolio
Mean 0.000233 0.000301 �0.000219 0.000171Variance 6.50 ð 10�6 1.410 ð 10�5 6.210 ð 10�6 3.691 ð 10�6
Skewness 0.0132 �1.063ŁŁ �0.5663ŁŁ �0.291ŁŁKurtosis 3.37ŁŁ 14.654ŁŁ 18.369ŁŁ 4.446ŁŁBera–Jarque Statistic 2300ŁŁ 44400ŁŁ 68700ŁŁ 4080ŁŁ
Notes : The Bera–Jarque statistic is distributed asymptotically as a �2(2) under the null of normality. Ł and ŁŁ indicate significance at the5% and 1% levels respectively.
while the government bond index returns is the least. The benefits from diversification, in terms ofa substantial reduction in variability, are clear, since the variance for the equally weighted portfolioreturns is almost half that of the least volatile component. Also, as one might anticipate, the seriesare all strongly non-normal. All are leptokurtic, while the FTSE All-Share and commodities seriesare also significantly skewed to the left.
FORECASTING VOLATILITY
Construction of forecasts and notationThe total sample of 4865 observations is split into two parts: the first 1250 observations (approx-imately 5 years of daily trading data) are used for estimation of the parameters of the model, andthen one-, two-,. . ., twenty-step-ahead forecasts are calculated. The multi-step-ahead forecasts arethen aggregated to form forecasts of volatility over the next 5, 10, and 20 days. We can thus write5
�2t,N D
N∑nD1
�2t,tCn �1�
where �2t,N denotes the time t aggregated forecast for the next N steps, and �2
t,tCn denotes then-step-ahead forecast made at time t.
In contrast to much previous research in this area, these are not one-, two-, three-. . . twenty-step-ahead forecasts, but rather we aggregate the forecasts for the next 5, 10, and 20 days. Aggregatedforecasts will be the ones of interest to financial market practitioners and risk managers, when theyhave investment horizons longer than one day; they will not be particularly interested in multi-step-ahead one-day volatility forecasts, such as the volatility forecast for day t C 20 made on day t. Thesample is then rolled forward by removing the first observation of the sample and adding one tothe end, and another set of forecasts of the next twenty days’ volatilities is made, and aggregated.This ‘recursive’ modelling and forecasting procedure is repeated until a forecast for observation4865 has been made using data available at time 4845. Computation of forecasts using a rollingwindow of data should ensure that the forecasts are made using models whose parameters have
5 This step is permissible since the variances are additive over time. Another possibility would be to multiply the one-step-ahead forecast bythe desired horizon using an equivalent of the ‘square root of time’ rule, so that, for example, the volatility forecast over the next 20 daysis 20 times the forecast for tomorrow. However, our approach is likely to be superior, since it employs more information while implicitextrapolation of one-step forecasts could be inappropriate for a mean-reverting series.
Copyright 2003 John Wiley & Sons, Ltd. J. Forecast. 22, 1–22 (2003)
Volatility Forecasting for Risk Management 5
been estimated using a sufficient span of time, while not incorporating such old vintages that thedata may no longer relevant in the context of an evolving financial market.
Forecasting modelsAlmost all of the forecasting models employed in this study are not new, rather it is the evaluationof the models which is novel. Hence the model descriptions are brief and presented in Table II,with �2
f,tCnj�t denoting the n-step-ahead (n D 1, 2, . . . , 20) forecast for the conditional varianceupon information available at time t, where t runs from observation 1250 to 4845. With onepossible exception, the model equations in Table II are self-explanatory, and readers are referred toBollerslev et al. (1992), Brailsford and Faff (1996), or Brooks (1998), and the references therein,for a more detailed treatment.
The only model which perhaps requires further explanation is the multivariate GARCH model,which has not been employed in previous studies of volatility forecast performance. The particularparameterization used here is of the diagonal VEC form due to Bollerslev, Engle, and Wooldridge(1988), where each element of the conditional variance covariance matrix hjk,t depends on pastvalues of itself and past values of εj,tε0
j,t, which may be written
vec�HtC1� D htC1 D C0 C A1vec�εtε0t� C B1ht �2�
where vec denotes the column stacking operator, A1 and B1 are restricted to be diagonal. Theparameterization for HtC1 conditional upon the information set allows each element of the condi-tional variance–covariance matrix to depend on lags of the squares and of the cross products ofthe elements of εtC1 as well as lags of the elements of HtC1.
EVALUATING VOLATILITY FORECASTS
Standard loss functionsThree criteria are used here to evaluate the accuracy of the forecasts: mean squared error (MSE),mean absolute error (MAE), and proportion of over-predictions. Mean squared error provides aquadratic loss function which disproportionately weights large forecast errors more heavily relativeto mean absolute error, and hence the former may be particularly useful in forecasting situationswhen large forecast errors are disproportionately more serious than small errors. The proportionof over-predictions should give a rough indication of the average direction of the forecast error(compared with the two previous measures which only give some measure of the average size) andwhether the models are persistently over- or under-predicting the ‘true’ value of volatility. Hencethis measure gives an approximate guide as to whether the forecasts are biased.
But what is volatility?Unlike financial asset returns, volatilities are not directly observable from the market. Consequently,when attempting to benchmark the accuracy of volatility forecasting models, researchers are nec-essarily required to make an auxiliary assumption about how the ex post or realized volatilities arecalculated. The vast majority of existing studies, including those listed in the introduction to thispaper, use squared returns of the frequency of the data and analysis, as the measure of realizedvolatility. For example, studies using daily data would assume that the ‘correct’ volatility number
Copyright 2003 John Wiley & Sons, Ltd. J. Forecast. 22, 1–22 (2003)
6 C. Brooks and G. Persand
Table II. Description of models used for forecasting
Model Acronym Equations for model Equation
1. Random walk in volatility RW �2f,tCn D �2
t (2)
2. Long-term mean LTM �2f,tCn D 1
1250
t∑jDt�1249
�2t�j (3)
3. Short-term moving average MA5 �2f,tCn D 1
5
4∑jD0
�2t�j (4)
4. Long-term moving average MA100 �2f,tCn D 1
100
99∑jD0
�2t�j (5)
5. Linear regression with one lag AR1 �2f,tCn D ˛0 C ˛1�
2t C εt (6)
6. Linear regression with AIC lags ARAIC �2f,tCn D ˇ0 C
p�1∑jD0
ˇj�2t�j C εt (7)
7. GARCH(1,1) GAR rtC1 D � C εtC1, εtC1 ¾ N�0, �2tC1�, (8)
�2f,tCn D �0 C ϕ1ε
2t C �2�
2t
8. GJR(1,1) GJR �2f,tCn D υ0 C υ1ε
2t C υ2�
2t C υ3S
�t ε2
t (9)S�
t D 1 for εt � 0 and 0 otherwise
9. EGARCH(1,1) EGAR log��2f,tCn� D ω1 C ω2 log��2
t � C ω3εt√�2
t
(10)
Cω4
j εt j√
�2t
�√
2
�
10. Long exponentially weightedmoving average
EMA5 �2f,tCn D �1 � 1�
5∑tD1
t�11 �rt � r� (11)
11. Short exponentially weightedmoving average
EMA100 �2f,tCn D �1 � 1�
100∑tD1
t�11 �rt � r� (12)
12. GARCH with t-distributed errors GART rtC1 D � C εtC1, εtC1 ¾ tk�0, �2tC1�, (13)
�2f,tCn D �0 C ϕ1ε
2t C �2�
2t
13. Multivariate GARCH MGAR See text for model description —
Notes : Forecast equations are given for n D 1 step ahead, and recursions can easily be computed from these for the 2, 3, . . . , 20 step-aheadforecasts. The model order p for ARAIC is determined individually for each forecast iteration by the minimization of Akaike’s informationcriterion, with maximal lag 5. All model parameters are estimated using quasi-maximum likelihood. The exponentially weighted movingaverage coefficients ( i) are chosen to produce the best fit by minimizing the sum of the squared in-sample forecast errors.
on day t is r2t , and it is this value that would be used as an input to the mean squared error calcu-
lation, or as the dependent variable in a Fair–Schiller (1990)-type regression of actual volatilitieson their forecasted values.
Whilst this method is simple and intuitively plausible, Andersen and Bollerslev (1998, hereafterAB) suggest that ‘same-frequency’ squared returns are an unbiased but extremely noisy measureof the latent volatility factor which underlies financial asset return movements. AB show that a
Copyright 2003 John Wiley & Sons, Ltd. J. Forecast. 22, 1–22 (2003)
Volatility Forecasting for Risk Management 7
much better approximation to the latent volatility factor can be obtained by summing the squares ofhigher frequency returns. For example, a superior estimate of volatility on day t to r2
t is given by
r2Łt D
m∑jD1
r2t�1C�j/m� �3�
where m is an intra-day sampling frequency, such as 8 for hourly data.6
Unfortunately, for many applications, the usefulness of this method is limited by the lack ofavailability of a sufficiently long span of higher-frequency returns. In the present paper, however,our analysis focuses upon daily, weekly, bi-weekly, and monthly forecasts. For the latter threehorizons, two methods of calculating ex post volatility are available, both of which are employedin this study. The first of these ex post measures, which may usefully be termed the traditionalmeasure, is to use weekly, bi-weekly or monthly squared returns.7 The second method, would be totake the daily returns, square them, and sum them over the relevant (5-, 10-, or 20-day) horizon.8
As AB note, it is not necessarily the case that the two ex post measures will give the same modelrankings, let alone the same values of the error measures. Thus a comparison of model rankingsunder the two approaches is a relevant question for research, which this paper makes the firstattempt to address.
Value at Risk calculationGiven the voluminous literature which almost unquestioningly evaluates volatility forecasts usingstandard loss functions, three sensible questions to ask are first, what are volatility forecastsuseful for, second, what is an appropriate loss function given this usage, and finally, will alter-native loss functions lead to approval of the same or similar models? Some answers to the firstof these questions are provided in the introduction to this paper. One use of volatility predic-tions, which has grown substantially in importance over recent years, is as an input to financialrisk management. In this paper, we thus employ a relevant ‘risk management’ loss function,which is based upon the calculation of an institution’s value at risk, as defined above in thefirst section. Specifically, we calculate VaR for three individual assets by calculating the follow-ing quantity:
VARit�N, ˛� D [Fi
t,N]�1( ˛
100
)�4�
where VARit is the Value at Risk for a given asset at time t, determined from model i (where
i D 1, 13 are the models as defined above), N is the investment horizon, [Fit,N]�1 is a cumulative
distribution function (cdf) and ˛ is a percentage significance level. The cdf employed in this paperis that of a normal distribution.
A limiting assumption of the analysis in many empirical papers in risk management is the stan-dard assumption of normality, for it is well known that asset returns are not Gaussian. However,the normal approximation is extremely widely used in the risk management field. Fat-tailed returndistributions will lead the delta-normal model to understate the true value at risk (see Jorion, 1996or Huisman et al., 1998). For example, a 5% daily loss is observed to occur approximately once
6 Assuming, of course, that 8 hourly observations are available from the financial market concerned.7 So, for example, the volatility for weekly returns would be given by r2
t D [ln�Pt/Pt�5�]2.8 Obviously for the 1-day horizon, both methods will yield the same ex post measure.
Copyright 2003 John Wiley & Sons, Ltd. J. Forecast. 22, 1–22 (2003)
8 C. Brooks and G. Persand
every two years, while if returns were normally distributed, such a change would be expectedonly once every 1000 years (Johansen and Sornette, 1999). A number of methods to incorpo-rate the fat tails have been proposed, most importantly the use of extreme value distributions forreturns (e.g. Embrechts et al., 1999). However, we continue to employ the normality assumptionsince other distributional approaches usually do not directly employ a volatility estimate. There-fore our purpose of comparing between volatility forecasts when used for risk management wouldbe lost.
We employ both the 1% and 5% levels of significance. The former level has been selected bythe Basle Committee (1996) as the focus of attention, although the first percentile of a distributionis more difficult to estimate than the fifth, and thus the latter is the quantity which many securitiesfirms wish to employ (see JP Morgan, 1996). The VaR corresponding to 5% may be defined asthat amount of capital, expressed as a percentage of the initial value of the position, which will berequired to cover 95% of probable losses. In the case of the normal distribution, this quantity maybe calculated as
VARit�N, 5%� D 1.645�i
t,N �5�
where �it,N is the square root of the conditional variance forecast, made at time t for forecast horizon
N (N D 1, 5, 10, 20). We thus forecast volatility for some future period (t, N) and hence we calculatethe amount of capital required to cover expected losses on 95% or 99% of the investment horizons.The 95% confidence level is employed by the popular RiskMetrics risk measurement software,while the regulators require capital to cover 99% of losses.9
The calculation of the value at risk estimates for the individual assets is achieved by followingthe steps outlined above. In the case of the portfolio, however, for all forecasting models except themultivariate GARCH (that is, models 1–12 in Table II), we employ a method known as the ‘fullvaluation approach’. This simply involves the aggregation of the components and the calculationof the portfolio return at each point in time. In this case, the resulting portfolio return series ismodelled in the same way as the individual component assets.
An alternative approach is known as the ‘volatilities and correlations’ method, which has beenpopularized by JP Morgan (1996). Here, the portfolio value at risk is estimated using the volatilitiesof the individual assets which form the MCRR, and the correlations between their returns. Theportfolio value at risk may be written
MCRRP D
√√√√√√a2MCRR2
A C b2MCRR2B C c2MCRR2
CC2ab�ABMCRRAMCRRB
C2ac�ACMCRRAMCRRC
C2bc�BCMCRRBMCRRC
�6�
where A, B, and C denote the bond, stock and commodities series respectively, and a D b D c D1/3. We adopt this approach when using the multivariate GARCH model, but instead of usingthe time-invariant volatility and correlation estimates, we instead use the relevant forecasts of theconditional variances and covariances from the MGARCH model in (6) to derive the VaR.
9 In fact, the 99% VaR is multiplied by a ‘scaling factor’, which is usually 3, so that the actual coverage rate is considerably higher than99%. We do not employ the regulatory scaling factor in our analysis, so as to focus upon forecast adequacy. Multiplying the estimated VaRby 3 has the effect of rendering the forecasted VaRs virtually indistinguishable from one another, since the implied coverage rate is nowmore than 99.99%.
Copyright 2003 John Wiley & Sons, Ltd. J. Forecast. 22, 1–22 (2003)
Volatility Forecasting for Risk Management 9
Risk management-based forecast evaluationsIn this paper, we employ three methods for determining the adequacy of the volatility forecaststhat are used as an input to the value at risk calculation. All methods essentially require thecalculation of VaR, and then assuming that the securities firm had employed this much capital, themethods track the actual realized losses during an out-of-sample period. The simplest approach todetermining model adequacy in the risk management framework is to calculate the time until firstfailure (TUFF), defined as the first observation in the hold-out sample where the capital held isinsufficient to absorb that period’s loss, and derived as follows. Following Kupiec (1995), let pdenote the realized probability of observing the first failure of the model in period V, and lettingQR be a random variable that denotes the number of observations until the first failure is recorded,then we may write
Pr� QR D V� D p�1 � p�V�1 �7�
Then QR has a geometric distribution with an expected value of 1/p. This quantity can be interpretedas the expected number of observations until the first failure is observed. In the cases of interestin this paper, if the actual proportion of failures were 5% and 1% respectively, then the time untilfirst failure would be 100 and 20 steps respectively. If we now let p* denote the probability offailure under the null hypothesis, then the following likelihood ratio test can be established:
TUFF �V, pŁ� D �2 log[p Ł �1 � pŁ�V�1] C 2 log
[1
V
(1 � 1
V
)V�1]
�8�
which is �2(1) under the null. Given the appropriate critical value, it is possible to derive a 95%confidence interval for TUFF of (6,439) for the 1% VaR and (�,87) for the 5% VaR.10 Theconfidence intervals can be interpreted as follows. If VaR determined using a 1% significance levelfails before the 6th observation, we can reject at the 5% level the null hypothesis that the model isadequate to cover losses on 99% of occasions. On the other hand, if the actual TUFF is greater than439, then we would conclude that the model was leading to too high a value at risk, and thereforethat the model was not failing as quickly as would be expected given the nominal 1% probabilityof failure.
It is perhaps worth noting that it is desirable from the point of view of the bank or securitiesfirm concerned, for the calculated value at risk to be neither too large nor too small. A value atrisk set too low could imply that the bank does not have sufficient capital to cover future losses,leading at best to regulatory scrutiny, and an increase in the scaling factor (resulting in a substantialincrease in the capital requirement), and at worst to financial distress and possible company failure.Conversely, a VaR set too high, so that it covers more than the nominal percentage of horizons(e.g. an estimated 5% daily VaR which is actually sufficient to cover 99.9% of the out of sampleperiods), probably implies that the firm is tying up too much of its capital unnecessarily in anunprofitable fashion.11
Whilst intuitive and simple to calculate, TUFF has obvious flaws as an evaluation metric. First,it is clearly not using much information from the sample, since all observations after the firstfailure are ignored, resulting in the test being over-sized. Thus, if the start of the out-of-sample
10 It is not possible to establish a lower limit for the 5% VaR interval.11 Particularly in view of the regulatory scaling factor, which multiplies the firm’s own VaR estimate by at least 3.
Copyright 2003 John Wiley & Sons, Ltd. J. Forecast. 22, 1–22 (2003)
10 C. Brooks and G. Persand
period occurs at a time of exceptional market turbulence, a model which may have been perfectlyadequate for the rest of the sample and incurring no further failures, would be rejected. Second,the TUFF statistic consequently has low power to reject models which are not adequate—thisis clearly evidenced by the wide confidence intervals for TUFF presented above. For example, a99% nominal coverage rate is expected to result in first failure at observation 100, but even if anexceedence of the VaR is recorded as early as observation 7, we cannot reject the underlying modelat the 1% level; thus TUFF will have low ability to discriminate between volatility forecasts fromdifferent models.
Another simple method for determining model adequacy within the risk management frameworkis simply to calculate the percentage of times that the calculated VaR is insufficient to coverthe actual losses, during the rolling out-of-sample period. A good model would be one whoseproportion of out-of-sample exceedences is close to the nominal value of (one minus coverageprobability)% assumed (5% or 1%). We can also formulate a likelihood ratio test for the proportionof failures, in similar vein to (19) above. The probability of observing x failures in an actual sampleof independent observations of size K will be distributed binomially, leading to the following teststatistic distributed �2(1) under the null:
UCF�K, x, pŁ� D �2 log[�1 � pŁ�K�x�pŁ�x] C 2 log[(
1 �( x
K
))K�x ( x
K
)x]
�9�
with notation as above. For ease of interpretation of the results, models are also ranked in thefollowing way. We assume that any model which has a percentage of exceedences in the rolling hold-out sample which is greater than the nominal threshold should be rejected as inadequate. Therefore,the lowest ranking models (classified as worst) are those which have the highest percentage offailures greater than the nominal value.
When these models have been exhausted, we assume further that any model which generates farfewer exceedencess than the expected number is less desirable than a model which generates closerto the nominal number. Thus the best models under this loss function are those which generate lessthan, but closest to, the assumed coverage rate.12
RESULTS
Statistical evaluation criteriaThe results for the volatility forecasts under standard statistical evaluation methods (percentage ofover-predictions, mean squared error, and mean absolute error) are presented in Tables III to VIfor the government bond, FTA All-Share, commodities and portfolio series respectively.
Considering first the one-step-ahead (1-day) forecast horizon, a number of important featuresemerge. As one might anticipate, the random walk in volatility model produces roughly equalnumbers of over-and under-predictions of realized volatility measured by the squared daily returns.On the other hand, all models over-predict volatility on average 70% of the time, except for thetwo EWMAs which over-predict more frequently than they under-predict. In all other respects, therandom walk in volatility produces uniformly poor forecasts.
12 Of course, this could be replaced by a simpler symmetric or any other loss function if the user desired.
Copyright 2003 John Wiley & Sons, Ltd. J. Forecast. 22, 1–22 (2003)
Volatility Forecasting for Risk Management 11
Tabl
eII
I.St
atis
tical
loss
func
tions
for
gove
rnm
ent
bond
(sam
efr
eque
ncy
squa
red
retu
rns
asex
post
mea
sure
)
Step
s1
510
20
Mod
els
%O
PR
ank
MSE
Ran
kM
AE
Ran
k%
OP
Ran
kM
SER
ank
MA
ER
ank
%O
PR
ank
MSE
Ran
kM
AE
Ran
k%
OP
Ran
kM
SER
ank
MA
ER
ank
RW
49.8
81
0.03
612
0.08
212
45.3
12
0.92
713
0.42
513
43.0
93
3.14
613
0.85
413
41.5
95
14.2
513
1.90
313
LTM
76.9
713
0.02
01D
0.06
99D
70.8
211
0.48
53D
0.33
710
69.3
511
1.35
73
0.67
36D
67.1
210
6.36
21D
1.49
46
MA
567
.01
30.
021
80.
068
7D59
.36
30.
545
90.
352
1157
.02
41.
693
100.
717
1155
.74
37.
923
101.
617
12M
A10
073
.63
60.
020
1D0.
066
3D67
.23
90.
485
3D0.
329
5D64
.34
91.
406
50.
673
6D62
.42
76.
777
41.
515
9A
R1
76.4
411
D0.
020
1D0.
069
9D70
.90
120.
484
20.
336
969
.37
121.
354
20.
672
567
.15
116.
362
1D1.
493
5A
RA
IC75
.22
100.
020
1D0.
068
7D69
.76
100.
481
10.
334
7D68
.73
101.
349
10.
668
466
.98
96.
393
31.
495
7DG
AR
73.8
57
0.02
01D
0.06
63D
65.9
57
0.48
96
0.32
74
60.8
95
1.44
04
0.66
13
57.3
34
7.07
16
1.48
24
GJR
76.4
411
D0.
024
10D
0.07
611
62.7
85
0.54
710
0.34
32
54.3
51
1.56
58
0.67
48D
46.5
72
7.61
88
1.48
02
EG
AR
69.9
65
0.17
013
0.13
513
48.0
11
0.79
612
0.41
212
37.8
36
2.00
012
0.77
212
28.7
612
8.32
411
1.64
010
EM
A5
43.4
82
0.02
29
0.05
61
37.8
94
0.54
28
0.30
21
35.8
68
1.62
89
0.62
71
34.3
88
7.87
09
1.44
01
EM
A10
031
.77
40.
024
10D
0.06
12
26.8
713
0.59
711
0.33
47D
25.7
613
1.86
011
0.70
210
24.5
313
8.76
512
1.57
611
GA
RT
74.1
68
0.02
01D
0.06
63D
67.2
08
0.48
53D
0.32
95D
63.4
27
1.43
06
0.67
48D
60.3
36
6.92
05
1.49
57D
MG
AR
74.8
39
0.02
01D
0.06
76
63.6
76
0.49
67
0.32
13
56.5
82
1.48
67
0.65
42
51.0
21
7.41
87
1.48
13
Tabl
eIV
.St
atis
tical
loss
func
tions
for
equi
ties
(sam
efr
eque
ncy
squa
red
retu
rns
asex
post
mea
sure
)
Step
s1
510
20
Mod
els
%O
PR
ank
MSE
Ran
kM
AE
Ran
k%
OP
Ran
kM
SER
ank
MA
ER
ank
%O
PR
ank
MSE
Ran
kM
AE
Ran
k%
OP
Ran
kM
SER
ank
MA
ER
ank
RW
50.2
11
0.42
09D
0.17
511
45.5
41
18.2
613
1.03
212
43.1
46
103.
213
2.30
813
41.0
36
440.
213
5.34
413
LTM
75.1
913
0.41
57D
0.15
69D
69.0
112
15.0
26
0.90
68D
65.9
012
78.8
73
1.92
59
62.3
99
313.
34
4.44
29
MA
563
.95
30.
389
50.
148
357
.89
415
.26
90.
904
754
.88
479
.81
81.
913
850
.38
131
5.9
54.
333
4M
A10
069
.37
60.
412
60.
156
9D63
.03
915
.07
70.
906
8D60
.25
779
.38
41.
977
1156
.44
531
3.0
2D4.
352
5A
R1
71.8
511
0.38
74
0.15
15
67.8
711
14.5
71
0.88
65
65.8
111
78.2
11
1.90
87
62.2
08
313.
02D
4.43
28
AR
AIC
70.3
28
0.41
57D
0.15
37
67.0
110
14.9
74
0.89
06
65.4
810
78.7
72
1.90
76
62.1
17
312.
71
4.42
77
GA
R70
.68
90.
342
10.
142
261
.25
814
.93
30.
842
355
.36
579
.47
61.
808
344
.70
431
8.9
74.
275
1G
JR72
.05
120.
420
9D0.
149
457
.52
315
.13
80.
830
146
.98
180
.05
91.
796
234
.94
1032
3.8
104.
363
6E
GA
R66
.15
40.
796
120.
274
1245
.40
215
.54
100.
929
1136
.97
880
.81
101.
984
1230
.63
1232
3.5
94.
610
12E
MA
543
.98
22.
859
130.
443
1339
.50
617
.38
121.
147
1335
.10
980
.89
111.
897
532
.63
1132
4.1
114.
517
10E
MA
100
32.1
85
0.47
511
0.15
58
29.3
513
15.9
111
0.92
210
26.9
813
81.3
712
1.94
310
24.5
613
325.
712
4.53
311
GA
RT
70.1
87
0.34
32
0.13
91
60.8
37
14.9
02
0.83
42
54.7
43
79.5
27
1.79
31
44.9
83
319.
88
4.28
62
MG
AR
71.7
410
0.37
03
0.15
26
60.2
05
15.0
15
0.85
74
54.2
42
79.4
25
1.85
04
45.7
02
317.
26
4.32
13
Copyright 2003 John Wiley & Sons, Ltd. J. Forecast. 22, 1–22 (2003)
12 C. Brooks and G. Persand
Tabl
eV
.St
atis
tical
loss
func
tions
for
com
mod
itie
s(s
ame
freq
uenc
ysq
uare
dre
turn
sas
expo
stm
easu
re)
Step
s1
510
20
Mod
els
%O
PR
ank
MSE
Ran
kM
AE
Ran
k%
OP
Ran
kM
SER
ank
MA
ER
ank
%O
PR
ank
MSE
Ran
kM
AE
Ran
k%
OP
Ran
kM
SER
ank
MA
ER
ank
RW
50.0
71
0.14
611
0.09
110
45.4
05
2.68
313
0.44
612
43.5
92
10.7
813
0.95
612
40.9
54
45.6
313
2.12
413
LTM
78.2
513
0.08
21D
0.07
62D
73.6
313
0.83
13
0.33
47
69.8
510
3.08
64
0.70
56D
66.3
18
16.1
74
1.57
55D
MA
567
.23
40.
095
7D0.
084
8D61
.53
71.
049
90.
369
959
.11
53.
736
90.
781
1155
.58
217
.51
71.
712
10M
A10
076
.97
80.
083
50.
079
5D72
.13
100.
835
40.
345
868
.71
93.
076
10.
721
864
.23
515
.89
11.
616
8A
R1
77.4
111
0.08
21D
0.07
51
73.3
511
0.83
02
0.33
35D
69.9
613
3.08
53
0.70
56D
66.2
06
16.1
63
1.57
67
AR
AIC
77.2
59
0.08
21D
0.07
62D
73.4
612
0.82
91
0.33
35D
69.9
011
D3.
084
20.
704
566
.23
716
.14
21.
575
5DG
AR
76.5
56
0.09
57D
0.08
17
54.9
46
0.89
48
0.29
63
41.9
83
3.34
97
0.64
12
29.0
710
17.8
58
1.53
41
GJR
77.7
512
0.10
89
0.08
48D
46.2
04
0.88
77
0.29
02
34.4
48
3.34
16
0.64
33
23.8
413
17.8
69
1.55
33
EG
AR
72.6
35
0.34
412
0.15
613
53.1
03
1.47
210
0.43
010
41.6
44
3.78
310
0.75
19
32.8
59
18.2
811
1.63
59
EM
A5
55.9
72
0.12
410
0.11
511
50.5
41
1.79
611
0.53
013
47.8
41
5.45
512
0.99
313
44.4
03
16.6
85
1.78
512
EM
A10
037
.36
30.
353
130.
125
1231
.15
91.
838
120.
434
1130
.10
11D
3.78
611
0.75
810
27.6
511
18.8
712
1.71
311
GA
RT
77.2
710
0.08
21D
0.07
74
51.4
12
0.86
55
0.28
41
39.0
37
3.37
48
0.63
91
26.2
912
18.0
010
1.54
62
MG
AR
76.8
07
0.09
36
0.07
95D
66.6
88
0.88
16
0.33
04
60.4
26
3.28
95
0.69
74
51.4
31
17.0
76
1.56
54
Tabl
eV
I.St
atis
tical
loss
func
tions
for
port
foli
o(s
ame
freq
uenc
ysq
uare
dre
turn
sas
expo
stm
easu
re)
Step
s1
510
20
Mod
els
%O
PR
ank
MSE
Ran
kM
AE
Ran
k%
OP
Ran
kM
SER
ank
MA
ER
ank
%O
PR
ank
MSE
Ran
kM
AE
Ran
k%
OP
Ran
kM
SER
ank
MA
ER
ank
RW
49.8
81
0.03
612
0.08
212
47.5
71
0.39
913
0.24
512
44.8
43
1.63
713
0.52
012
42.1
73
6.63
713
1.14
412
LTM
76.9
713
0.02
01D
0.06
99D
70.4
010
0.20
01
0.19
67
68.4
610
0.75
61
0.40
86
62.8
77
2.92
71
0.90
04D
MA
567
.01
30.
021
80.
068
7D60
.81
40.
259
100.
209
1158
.03
40.
955
100.
441
1053
.21
13.
420
80.
969
9M
A10
073
.63
60.
020
1D0.
066
3D66
.98
70.
201
20.
193
663
.89
70.
764
20.
405
558
.72
43.
035
40.
914
8A
R1
76.4
411
D0.
020
1D0.
069
9D69
.65
90.
212
50.
197
8D68
.26
90.
772
30.
409
762
.92
82.
950
30.
902
6A
RA
IC75
.22
100.
020
1D0.
068
7D69
.40
80.
218
6D0.
197
8D68
.01
80.
777
40.
410
862
.70
62.
930
20.
900
4DG
AR
73.8
57
0.02
01D
0.06
63D
60.8
95
0.21
04
0.18
23D
51.6
31
0.82
37
0.38
74
40.3
95
3.39
97
0.89
43
GJR
76.4
411
D0.
024
10D
0.07
611
54.0
82
0.22
48
0.16
11
39.9
45
0.78
45
0.32
71
28.1
010
3.06
45
0.79
51
EG
AR
69.9
65
0.17
013
0.13
513
44.7
03
0.21
86D
0.17
12
36.1
66
0.86
48
0.38
02
29.7
99
3.58
110
0.90
77
EM
A5
43.4
82
0.02
29
0.05
61
28.4
311
0.23
19
0.19
15
25.4
011
0.90
09
0.41
59
22.0
312
3.70
011
0.97
110
EM
A10
031
.77
40.
024
10D
0.06
12
20.2
212
0.28
912
0.20
310
18.4
712
1.16
412
0.44
511
15.9
413
3.88
812
0.99
411
GA
RT
74.1
68
0.02
01D
0.06
63D
62.8
76
0.20
53
0.18
23D
54.2
72
0.80
26
0.38
33
43.3
92
3.33
46
0.88
52
MG
AR
74.8
39
0.02
01D
0.06
76
87.8
213
0.28
211
0.34
613
83.2
013
0.99
711
0.65
913
75.7
411
3.48
99
1.26
113
Copyright 2003 John Wiley & Sons, Ltd. J. Forecast. 22, 1–22 (2003)
Volatility Forecasting for Risk Management 13
No clear ‘winners’ emerge at the 1-day horizon, with different models being preferred for eachseries. MSE is clearly not a good discriminator at the top end, with many models ranked equallyas the best. MAE, on the other hand, selects EWMA models for the bond and portfolio series,while for the share and commodity series, the GJR and autoregressive volatility models are pre-ferred. In terms of the least accurate next-day forecasting models, the random walk in volatilityand EGARCH models emerge as the worst performers, followed by the EWMAs for commoditiesand shares, although the latter proved the most accurate for the other two series.
An extension of the forecast and investment horizon to the one (trading) week, two-week, and one-month range does not markedly alter the relative model rankings, although the broad disagreementbetween criteria for a given series and model is still apparent. For example, the autoregressivemodel, which ranks only seventh by MSE for the equities series at the one-day horizon, ranks firstwhen the investment horizon is extended to one month.
However, as Andersen and Bollerslev (1998) have shown, the use of low-frequency squaredreturns is often not a useful way to evaluate volatility forecasts, and it is quite possible that whensums of higher-frequency squared returns are used instead as the ex post volatility measure, notonly the values of the error measures but also the model rankings could change substantially.Thus for the 5-, 10-, and 20-day periods we also evaluate the forecast accuracies using the sum ofsquared daily returns. Results are presented for the bond, share, commodities, and portfolio series inTables VII to X respectively.13 Comparing the results for the low-frequency squared returns versusthe high-frequency sums of squared returns, we note first that the values of the error measures areas expected reduced considerably.14
The GARCH model with t-distributed errors now emerges as the clear winner, producing themost accurate forecasts according to MAE, for three of the four series (bonds, stocks, and theportfolio). Only for the commodities return series does GARCH-t perform poorly. For the latterseries, the long-term mean and autoregressive volatility models prove to be the best under bothsquared and absolute error measures. Interestingly, the worst models seem invariant to both the useof a same-frequency or higher-frequency ex post measure, and to whether the errors are squaredor the absolute values taken; a bad model appears to be a bad model whatever. Models which fitinto this category are the random walk in volatility, the exponential GARCH, and the exponentiallyweighted moving average model.
Risk management evaluation criteriaThe corresponding evaluations for the forecasts when used in a risk management context are givenin Tables XI to XVIII. Volatility forecasts can be employed for the production of 99% and 95%nominal coverage rates for the value at risk estimates. In other words, forecasts are generatedin respect of the amount of capital required to cover expected losses on 99% and 95% of daysrespectively. The results for these two sets of nominal coverage rates are provided in Tables XI toXIV and XV to XVIII respectively for the 1-day, 1-week, 2-week and 1-month horizons. Threestatistics are presented in each table—the time until first failure (TUFF), the proportion of failures(FT), and the test statistic associated with whether this proportion of failures is significantly higherthan the nominal rate (UCF). Also given are the model rankings according to FT and UCF15 asdescribed above.
13 Of course, the results for the one-step-ahead evaluations will be identical to those of Tables III to VI.14 Mean squared errors are reduced by roughly an order equivalent to the forecasting horizon, while absolute errors are reduced by a factorof around two for all horizons.15 The rankings according to FT and UCF will, of course, by definition be identical.
Copyright 2003 John Wiley & Sons, Ltd. J. Forecast. 22, 1–22 (2003)
14 C. Brooks and G. Persand
Tabl
eV
II.
Stat
istic
allo
ssfu
nctio
nsfo
rgo
vern
men
tbo
nd(s
umof
daily
squa
red
retu
rns
asex
post
mea
sure
)
Step
s1
510
20
Mod
els
%O
PR
ank
MSE
Ran
kM
AE
Ran
k%
OP
Ran
kM
SER
ank
MA
ER
ank
%O
PR
ank
MSE
Ran
kM
AE
Ran
k%
OP
Ran
kM
SER
ank
MA
ER
ank
RW
49.8
81
0.03
612
0.08
212
35.2
76
0.55
313
0.34
413
32.1
87
2.05
213
0.66
013
30.4
09
8.04
213
1.30
813
LTM
76.9
713
0.02
01D
0.06
99D
71.0
710
0.14
37
0.23
28
68.6
59
0.37
66
0.40
68
68.2
96
1.04
25
0.72
57
MA
567
.01
30.
021
80.
068
7D51
.18
10.
181
80.
231
746
.62
20.
561
100.
407
943
.26
31.
716
90.
788
9M
A10
073
.63
60.
020
1D0.
066
3D64
.17
50.
137
50.
214
461
.14
50.
359
40.
369
460
.67
41.
016
30.
667
3A
R1
76.4
411
D0.
020
1D0.
069
9D71
.10
110.
141
60.
230
668
.51
80.
372
50.
403
768
.32
71.
036
40.
722
6A
RA
IC75
.22
100.
020
1D0.
068
7D70
.38
90.
136
40.
222
568
.82
100.
354
30.
387
568
.35
81.
001
10.
705
5G
AR
73.8
57
0.02
01D
0.06
63D
62.4
54
0.13
02
0.20
12
56.0
53
0.34
72
0.34
11
47.4
31
1.07
76
0.63
32
GJR
76.4
411
D0.
024
10D
0.07
611
57.4
12
0.18
610
0.23
39
43.8
74
0.47
28
0.40
16
29.7
410
1.49
58
0.76
18
EG
AR
69.9
65
0.17
013
0.13
513
32.1
08
0.42
112
0.33
312
20.4
711
0.90
812
0.60
312
15.0
811
2.69
012
1.18
012
EM
A5
43.4
82
0.02
29
0.05
61
18.8
912
0.18
49
0.23
710
14.6
312
0.54
49
0.45
810
12.7
412
1.73
710
0.89
610
EM
A10
031
.77
40.
024
10D
0.06
12
9.04
013
0.22
811
0.27
711
6.89
813
0.73
111
0.54
911
6.37
013
2.53
511
1.09
311
GA
RT
74.1
68
0.02
01D
0.06
63D
65.3
17
0.12
81
0.20
43
61.2
56
0.33
61
0.34
52
55.1
62
1.00
82
0.62
81
MG
AR
74.8
39
0.02
01D
0.06
76
58.8
03
0.13
53
0.20
01
47.9
01
0.37
77
0.34
73
35.7
45
1.20
47
0.67
14
Tabl
eV
III.
Stat
istic
allo
ssfu
nctio
nsfo
req
uiti
es(s
umof
daily
squa
red
retu
rns
asex
post
mea
sure
)
Step
s1
510
20
Mod
els
%O
PR
ank
MSE
Ran
kM
AE
Ran
k%
OP
Ran
kM
SER
ank
MA
ER
ank
%O
PR
ank
MSE
Ran
kM
AE
Ran
k%
OP
Ran
kM
SER
ank
MA
ER
ank
RW
50.2
11
0.42
09D
0.17
511
36.4
76
9.77
513
0.77
412
34.7
26
40.4
013
1.55
013
33.0
24
161.
313
3.14
713
LTM
75.1
913
0.41
57D
0.15
69D
72.1
611
4.83
37
0.58
29
70.7
77
13.0
06
1.06
59
71.9
98
34.2
47
2.00
18
MA
563
.95
30.
389
50.
148
351
.18
14.
898
90.
522
546
.43
413
.17
80.
937
541
.47
133
.72
31.
734
4M
A10
069
.37
60.
412
60.
156
9D64
.03
74.
891
80.
541
762
.75
513
.28
90.
991
758
.67
232
.30
11.
657
3A
R1
71.8
511
0.38
74
0.15
15
71.1
08
4.61
41
0.55
28
70.6
38
12.6
21
1.02
98
71.5
76
33.8
44
1.96
17
AR
AIC
70.3
28
0.41
57D
0.15
37
71.2
99
4.72
94
0.53
16
71.8
210
12.6
93
0.98
56
71.7
97
33.6
72
1.89
96
GA
R70
.68
90.
342
10.
142
259
.39
54.
677
30.
450
247
.96
112
.67
20.
789
126
.90
934
.12
51.
587
1G
JR72
.05
120.
420
9D0.
149
451
.41
24.
773
60.
464
329
.93
913
.10
70.
888
46.
6512
36.7
99
2.03
19
EG
AR
66.1
54
0.79
612
0.27
412
27.9
610
5.25
710
0.67
811
20.7
511
14.0
211
1.27
512
17.5
810
37.9
311
2.49
512
EM
A5
43.9
82
2.85
913
0.44
313
20.8
312
7.32
812
0.86
713
14.0
812
13.8
810
1.14
510
12.0
211
37.8
210
2.27
010
EM
A10
032
.18
50.
475
110.
155
810
.35
135.
540
110.
663
107.
455
1314
.06
121.
204
116.
175
1338
.01
122.
369
11G
AR
T70
.18
70.
343
20.
139
159
.05
44.
674
20.
443
147
.34
312
.79
50.
792
228
.93
534
.35
81.
603
2M
GA
R71
.74
100.
370
30.
152
658
.47
34.
743
50.
475
447
.84
212
.74
40.
867
336
.89
334
.16
61.
741
5
Copyright 2003 John Wiley & Sons, Ltd. J. Forecast. 22, 1–22 (2003)
Volatility Forecasting for Risk Management 15
Tabl
eIX
.St
atis
tical
loss
func
tions
for
itco
mm
oditi
es(s
umof
daily
squa
red
retu
rns
asex
post
mea
sure
)
Step
s1
510
20
Mod
els
%O
PR
ank
MSE
Ran
kM
AE
Ran
k%
OP
Ran
kM
SER
ank
MA
ER
ank
%O
PR
ank
MSE
Ran
kM
AE
Ran
k%
OP
Ran
kM
SER
ank
MA
ER
ank
RW
50.0
71
0.14
611
0.09
110
34.5
57
2.42
513
0.42
312
30.5
24
8.99
913
0.82
812
27.4
07
34.2
113
1.63
013
LTM
78.2
513
0.08
21D
0.07
62D
73.8
312
0.50
81D
0.26
85D
71.1
37
1.05
91
0.46
63
68.5
74
2.29
33
0.80
23
MA
567
.23
40.
095
7D0.
084
8D51
.13
10.
735
90.
320
945
.31
21.
739
100.
587
940
.72
13.
746
91.
044
6M
A10
076
.97
80.
083
50.
079
5D68
.23
80.
523
40.
289
864
.70
31.
120
40.
508
660
.36
22.
557
40.
890
5A
R1
77.4
111
0.08
21D
0.07
51
73.2
79D
0.50
81D
0.26
74
71.2
48
1.06
02D
0.46
51D
68.6
55
2.29
22
0.80
12
AR
AIC
77.2
59
0.08
21D
0.07
62D
73.2
79D
0.50
81D
0.26
85D
71.1
06
1.06
02D
0.46
51D
68.7
66
2.29
11
0.80
01
GA
R76
.55
60.
095
7D0.
081
741
.25
20.
586
80.
255
216
.63
101.
306
60.
501
53.
143
123.
378
61.
086
7G
JR77
.75
120.
108
90.
084
8D26
.70
110.
575
6D0.
260
37.
677
131.
312
70.
542
82.
337
133.
472
81.
152
9E
GA
R72
.63
50.
344
120.
156
1336
.33
41.
156
100.
396
1019
.94
91.
758
110.
634
1011
.60
93.
902
101.
195
10E
MA
555
.97
20.
124
100.
115
1135
.13
51.
556
120.
519
1329
.46
53.
817
120.
891
1324
.42
84.
358
121.
329
12E
MA
100
37.3
63
0.35
313
0.12
512
14.4
713
1.51
011
0.42
111
11.2
712
1.68
89
0.65
511
10.1
310
4.20
811
1.25
011
GA
RT
77.2
710
0.08
21D
0.07
74
34.9
96
0.55
05
0.24
91
13.0
711
1.31
38
0.52
07
4.53
411
3.45
67
1.11
38
MG
AR
76.8
07
0.09
36
0.07
95D
61.3
13
0.57
56D
0.27
17
50.0
41
1.24
05
0.47
74
38.6
13
2.92
75
0.88
14
Tabl
eX
.St
atis
tical
loss
func
tions
for
port
foli
o(s
umof
daily
squa
red
retu
rns
asex
post
mea
sure
)
Step
s1
510
20
Mod
els
%O
PR
ank
MSE
Ran
kM
AE
Ran
k%
OP
Ran
kM
SER
ank
MA
ER
ank
%O
PR
ank
MSE
Ran
kM
AE
Ran
k%
OP
Ran
kM
SER
ank
MA
ER
ank
RW
49.8
81
0.01
29D
0.04
511
35.8
65
0.25
713
0.19
912
33.1
04
0.99
713
0.38
812
31.4
36
3.85
013
0.76
512
LTM
74.9
110
0.00
91D
0.03
87D
71.1
09
0.08
82
0.12
36
69.0
47D
0.20
81
0.21
26
68.1
05
0.49
42
0.36
83
MA
565
.09
30.
010
4D0.
039
951
.07
20.
136
100.
140
946
.62
10.
379
100.
250
743
.53
10.
837
90.
462
7M
A10
071
.54
50.
009
1D0.
037
4D64
.62
60.
090
30.
118
363
.17
30.
217
40.
206
360
.06
20.
543
40.
371
4A
R1
72.3
59
0.01
04D
0.03
87D
69.6
88
0.09
34
0.12
25
68.9
96
0.21
42
0.21
05
67.7
34
0.49
53
0.36
42
AR
AIC
71.9
36D
0.01
811
0.04
010
69.2
17
0.09
76
0.12
14
69.0
47D
0.21
63
0.20
94
67.3
73
0.49
01
0.36
21
GA
R72
.21
80.
010
4D0.
037
4D50
.77
10.
095
50.
109
232
.99
50.
242
60.
203
214
.35
90.
669
60.
439
6G
JR75
.05
110.
044
130.
046
1236
.41
40.
111
80.
137
822
.31
90.
275
80.
269
916
.36
80.
770
70.
538
8E
GA
R60
.33
20.
010
4D0.
033
117
.86
100.
100
70.
125
74.
089
120.
273
70.
261
80.
362
130.
788
80.
548
9E
MA
530
.71
40.
010
4D0.
035
28.
567
110.
112
90.
159
105.
953
110.
309
90.
317
105.
007
110.
917
100.
636
10E
MA
100
22.0
912
0.01
29D
0.03
63
4.36
713
0.16
311
0.17
511
3.28
213
0.53
112
0.35
211
2.97
612
0.96
411
0.65
511
GA
RT
71.9
36D
0.00
91D
0.03
74D
54.1
33
0.08
71
0.10
71
37.5
82
0.21
95
0.19
11
16.9
47
0.61
25
0.40
75
MG
AR
91.3
213
0.03
212
0.08
113
94.5
512
0.17
312
0.29
913
91.2
910
0.48
411
0.52
513
85.7
910
1.27
612
0.91
013
Copyright 2003 John Wiley & Sons, Ltd. J. Forecast. 22, 1–22 (2003)
16 C. Brooks and G. Persand
Tabl
eX
I.G
over
nmen
tbo
nd(r
isk
man
agem
ent
eval
uatio
n—
1%V
aR)
Step
s1
510
20
Mod
els
TU
FFFT
UC
FR
ank
TU
FFFT
UC
FR
ank
TU
FFFT
UC
FR
ank
TU
FFFT
UC
FR
ank
RW
312
.41
640.
712
111
.77
588.
612
110
.96
524.
610
19.
990
450.
111
LTM
331.
613
5.00
47D
396
0.96
70.
011
143
70.
974
0.01
11D
429
1.16
80.
424
1DM
A5
233.
255
50.3
39
983.
227
49.2
99
942.
726
31.9
48
33.
004
41.2
57
MA
100
231.
530
3.81
14D
221.
307
1.35
94
221.
335
1.60
35
429
1.47
43.
095
4A
R1
331.
613
5.00
47D
397
0.94
60.
047
243
70.
974
0.01
11D
429
1.16
80.
424
1DA
RA
IC33
1.53
03.
811
4D39
71.
196
0.57
13
437
0.94
60.
047
342
91.
196
0.57
13
GA
R23
1.50
23.
445
2D22
1.58
64.
590
626
71.
808
8.30
96
428
2.28
118
.99
6G
JR23
1.44
62.
762
122
2.78
233
.73
839
73.
922
77.4
99
145.
814
172.
99
EG
AR
121
3.39
455
.65
1020
8.73
435
9.3
106
12.1
361
7.9
123
14.5
282
1.2
12E
MA
515
11.2
154
4.2
1120
11.6
657
9.6
1121
11.4
956
6.3
114
9.84
743
9.7
10E
MA
100
318
.55
1195
131
18.0
511
4713
118
.11
1152
131
16.1
696
9.0
13G
AR
T23
1.50
23.
445
2D22
1.33
51.
603
543
41.
224
0.73
84
141.
558
4.19
35
MG
AR
231.
558
4.19
36
212.
031
12.9
07
212.
420
22.7
67
903.
727
69.1
68
Tabl
eX
II.
Equ
ities
(ris
km
anag
emen
tev
alua
tion
—1%
VaR
)
Step
s1
510
20
Mod
els
TU
FFFT
UC
FR
ank
TU
FFFT
UC
FR
ank
TU
FFFT
UC
FR
ank
TU
FFFT
UC
FR
ank
RW
911
.79
590.
812
3010
.99
526.
712
4410
.24
468.
910
649.
708
429.
410
LTM
444
1.69
76.
334
744
01.
752
7.29
32
713
1.50
23.
445
370
31.
975
11.6
83
MA
544
2.42
022
.76
912
23.
700
67.9
99
397
3.47
758
.93
711
73.
978
79.9
35
MA
100
231.
641
5.43
25D
302.
197
16.8
64
713
1.72
56.
806
470
32.
197
16.8
64
AR
123
1.64
15.
432
5D44
01.
780
7.79
43
713
1.44
62.
762
270
31.
947
11.0
82
AR
AIC
397
1.78
07.
794
844
01.
613
5.00
41
713
1.25
20.
925
170
31.
780
7.79
41
GA
R23
1.39
12.
148
213
22.
420
22.7
65
397
2.86
536
.48
611
84.
840
120.
86
GJR
231.
419
2.44
63
303.
366
54.5
78
126
4.59
010
8.4
910
97.
594
281.
89
EG
AR
163.
115
45.2
010
308.
651
353.
410
4413
.05
694.
012
6515
.83
938.
512
EM
A5
2310
.79
511.
611
3011
.13
537.
611
119
12.1
061
5.6
1111
410
.85
515.
911
EM
A10
016
17.1
610
6213
3017
.25
1070
1368
17.8
311
2513
108
16.3
398
4.5
13G
AR
T23
1.47
43.
095
413
22.
531
25.9
66
397
2.69
831
.06
511
85.
007
129.
37
MG
AR
231.
307
1.35
91
302.
587
27.6
27
121
3.53
361
.15
810
96.
203
195.
48
Copyright 2003 John Wiley & Sons, Ltd. J. Forecast. 22, 1–22 (2003)
Volatility Forecasting for Risk Management 17
Tabl
eX
III.
Com
mod
ities
(ris
km
anag
emen
tev
alua
tion
—1%
VaR
)
Step
s1
510
20
Mod
els
TU
FFFT
UC
FR
ank
TU
FFFT
UC
FR
ank
TU
FFFT
UC
FR
ank
TU
FFFT
UC
FR
ank
RW
115
.02
865.
612
116
.69
1018
1255
17.3
910
8312
4520
.17
1355
9LT
M1
1.69
76.
334
136
1.28
01.
132
1D67
1.72
56.
806
159
2.39
221
.99
1M
A5
364.
312
95.0
810
364.
256
92.4
96
615.
035
130.
75
537.
510
276.
45
MA
100
12.
114
14.8
38
361.
864
9.37
94
692.
142
15.4
94
355
2.61
528
.47
4A
R1
11.
864
9.37
92D
361.
280
1.13
21D
671.
752
7.29
32D
592.
420
22.7
62D
AR
AIC
11.
864
9.37
92D
361.
280
1.13
21D
671.
752
7.29
32D
592.
420
22.7
62D
GA
R1
2.00
312
.28
636
6.23
119
7.0
760
13.4
172
4.7
850
22.8
416
3010
GJR
361.
947
11.0
85
368.
679
355.
49
3616
.86
1033
1136
26.1
519
8812
EG
AR
363.
282
51.3
89
368.
790
363.
210
3614
.38
808.
99
3622
.06
1548
8E
MA
51
9.51
341
5.1
1136
10.8
551
5.9
1136
12.7
166
6.1
736
15.2
788
8.0
7E
MA
100
119
.69
1307
1336
23.3
916
8913
3624
.45
1802
1336
26.2
620
0113
GA
RT
11.
919
10.5
04
366.
843
233.
98
3614
.99
863.
110
3624
.42
1799
11M
GA
R36
2.03
112
.90
736
3.25
550
.33
536
5.48
015
4.4
636
8.73
435
9.3
6
Tabl
eX
IV.
Port
folio
(ris
km
anag
emen
tev
alua
tion
—1%
VaR
)
Step
s1
510
20
Mod
els
TU
FFFT
UC
FR
ank
TU
FFFT
UC
FR
ank
TU
FFFT
UC
FR
ank
TU
FFFT
UC
FR
ank
RW
1212
.24
626.
911
1211
.63
577.
411
3610
.74
507.
310
3610
.24
468.
99
LTM
374
1.14
00.
298
2D37
01.
113
0.19
32
367
0.94
60.
047
1D35
91.
140
0.29
81
MA
516
3.03
242
.22
975
3.17
147
.23
813
22.
893
37.4
26
117
3.22
749
.29
6M
A10
023
1.50
23.
445
837
01.
168
0.42
45
368
1.25
20.
925
535
91.
391
2.14
85
AR
123
1.33
51.
603
637
01.
140
0.29
83
367
0.94
60.
047
1D35
91.
168
0.42
42D
AR
AIC
361.
474
3.09
57
370
1.22
40.
738
436
70.
918
0.10
93
359
1.16
80.
424
2DG
AR
231.
140
0.29
82D
370
2.44
823
.55
736
53.
588
63.4
18
122
6.70
422
5.4
8G
JR37
41.
140
0.29
82D
307.
566
280.
010
4411
.27
548.
611
6016
.30
981.
912
EG
AR
233.
672
66.8
410
307.
510
276.
49
688.
039
311.
49
6810
.40
481.
610
EM
A5
1618
.19
1160
1229
16.8
310
3112
3615
.83
938.
512
116
15.9
194
6.1
11E
MA
100
325
.06
1869
1329
22.6
916
1513
3622
.73
1618
1359
20.9
514
3413
GA
RT
231.
140
0.29
82D
370
1.89
29.
934
636
52.
949
39.3
17
122
5.53
515
7.4
7M
GA
R37
40.
139
18.4
31
370
0.41
76.
862
136
70.
556
3.69
84
355
1.16
80.
424
2D
Copyright 2003 John Wiley & Sons, Ltd. J. Forecast. 22, 1–22 (2003)
18 C. Brooks and G. Persand
Tabl
eX
V.
Gov
ernm
ent
bond
(ris
km
anag
emen
tev
alua
tion
—5%
VaR
)
Step
s1
510
20
Mod
els
TU
FFFT
UC
FR
ank
TU
FFFT
UC
FR
ank
TU
FFFT
UC
FR
ank
TU
FFFT
UC
FR
ank
RW
315
.88
943.
611
115
.22
883.
011
114
.30
801.
610
112
.99
689.
410
LTM
233.
922
77.4
95D
223.
477
58.9
36
396
3.56
162
.28
442
83.
255
50.3
34
MA
53
6.62
022
0.3
91
6.98
224
2.6
91
6.06
418
7.3
73
6.00
818
4.1
7M
A10
022
4.72
911
5.2
121
4.50
610
4.3
221
4.00
681
.16
214
3.95
078
.70
2A
R1
233.
922
77.4
95D
223.
449
57.8
35
396
3.47
758
.93
542
83.
227
49.2
95
AR
AIC
233.
922
77.4
95D
223.
755
70.3
34
396
3.42
156
.74
642
83.
282
51.3
83
GA
R23
4.42
310
0.3
221
4.84
012
0.8
121
4.59
010
8.4
14
5.34
114
6.9
6G
JR23
3.95
078
.70
821
6.31
420
1.9
821
7.56
628
0.0
97
9.01
337
8.9
9E
GA
R22
6.84
323
3.9
102
13.2
971
5.2
102
16.5
510
0512
218
.03
1144
12E
MA
53
16.9
710
4412
116
.58
1008
121
16.0
595
8.9
113
14.5
582
3.6
11E
MA
100
223
.70
1722
131
22.7
316
1913
121
.92
1534
131
19.5
812
9713
GA
RT
224.
117
86.1
33D
204.
451
101.
73
213.
978
79.9
33
44.
423
100.
31
MG
AR
224.
117
86.1
33D
205.
257
142.
47
216.
147
192.
18
126.
926
239.
28
Tabl
eX
VI.
Equ
ities
(ris
km
anag
emen
tev
alua
tion
—5%
VaR
)
Step
s1
510
20
Mod
els
TU
FFFT
UC
FR
ank
TU
FFFT
UC
FR
ank
TU
FFFT
UC
FR
ank
TU
FFFT
UC
FR
ank
RW
915
.55
913.
211
3015
.08
870.
611
3013
.77
755.
710
6413
.27
712.
99D
LTM
233.
755
70.3
35
304.
701
113.
81
400
3.64
465
.69
112
23.
866
75.0
73
MA
516
6.31
420
1.9
974
7.73
329
0.9
974
7.01
024
4.4
611
37.
371
267.
45
MA
100
234.
339
96.3
82
305.
369
148.
44
126
5.09
013
3.6
411
85.
508
155.
94
AR
123
3.81
172
.69
430
4.61
810
9.7
240
03.
616
64.5
42
122
3.92
277
.49
2A
RA
IC23
4.39
599
.00
130
4.50
610
4.3
312
73.
449
57.8
33
118
3.95
078
.70
1G
AR
233.
588
63.4
16D
306.
064
187.
35
121
6.81
523
2.3
511
48.
818
365.
16
GJR
233.
588
63.4
16D
307.
650
285.
58
119
9.37
440
4.9
910
813
.27
712.
99D
EG
AR
156.
426
208.
610
3013
.38
722.
310
3017
.64
1107
1230
18.9
412
3412
EM
A5
1616
.41
992.
212
3015
.16
878.
112
4416
.13
966.
511
108
14.7
984
5.8
11E
MA
100
922
.70
1616
1330
22.1
715
6013
6821
.75
1517
1365
20.3
913
7813
GA
RT
233.
978
79.9
33
306.
092
188.
96
121
7.17
725
4.9
711
49.
013
378.
97
MG
AR
163.
477
58.9
38
306.
592
218.
67
447.
622
283.
78
108
10.4
348
3.8
8
Copyright 2003 John Wiley & Sons, Ltd. J. Forecast. 22, 1–22 (2003)
Volatility Forecasting for Risk Management 19
Tabl
eX
VII
.C
omm
oditi
es(r
isk
man
agem
ent
eval
uatio
n—
5%V
aR)
Step
s1
510
20
Mod
els
TU
FFFT
UC
FR
ank
TU
FFFT
UC
FR
ank
TU
FFFT
UC
FR
ank
TU
FFFT
UC
FR
ank
RW
119
.11
1250
121
21.5
314
9412
5522
.17
1560
1045
24.8
118
428
LTM
14.
061
83.6
38
364.
506
104.
33
364.
951
126.
41
366.
314
201.
91D
MA
51
8.48
434
1.8
1036
8.95
737
4.9
636
10.5
449
2.3
550
13.5
773
8.9
5M
A10
01
4.70
111
3.8
336
4.89
612
3.6
136
5.64
716
3.6
436
7.48
327
4.6
4A
R1
14.
618
109.
75
364.
562
107.
02
364.
784
117.
93
366.
370
205.
23
AR
AIC
14.
534
105.
76
364.
478
102.
94
364.
812
119.
42
366.
314
201.
91D
GA
R1
4.95
112
6.4
136
12.6
365
9.1
736
20.9
514
348
3630
.35
2469
10G
JR36
4.31
295
.08
736
15.8
393
8.5
1036
24.4
518
0312
3632
.16
2684
12E
GA
R1
6.78
723
0.5
936
14.7
484
0.8
936
21.2
814
689
3627
.82
2177
9E
MA
51
14.7
283
8.4
1136
17.3
310
7811
3618
.97
1236
736
22.0
315
467
EM
A10
01
24.8
418
4513
3629
.40
2359
1336
31.1
325
6113
3632
.38
2711
13G
AR
T1
4.67
311
2.4
436
13.5
773
8.9
836
22.5
616
0111
3631
.15
2565
11M
GA
R1
4.78
411
7.9
236
7.17
725
4.9
536
10.6
550
0.8
636
15.5
291
0.7
6
Tabl
eX
VII
I.Po
rtfo
lio(r
isk
man
agem
ent
eval
uatio
n—
5%V
aR)
Step
s1
510
20
Mod
els
TU
FFFT
UC
FR
ank
TU
FFFT
UC
FR
ank
TU
FFFT
UC
FR
ank
TU
FFFT
UC
FR
ank
RW
1215
.66
923.
311
1215
.27
888.
011
3614
.55
823.
610
3613
.38
722.
39
LTM
233.
922
77.4
96
342
3.50
560
.04
235
23.
060
43.2
13
342
2.86
536
.48
4M
A5
166.
871
235.
79
746.
843
233.
98
685.
925
179.
36
116
7.12
125
1.4
6M
A10
023
4.53
410
5.7
330
4.45
110
1.7
135
23.
644
65.6
91
342
4.11
786
.13
1A
R1
234.
590
108.
42
342
3.39
455
.65
435
23.
115
45.2
02
342
3.00
441
.25
3A
RA
IC23
4.86
812
2.2
134
23.
477
58.9
33
352
3.00
441
.25
434
23.
060
43.2
12
GA
R23
4.28
493
.78
574
5.67
516
5.1
769
7.87
230
0.2
811
611
.96
604.
38
GJR
233.
755
70.3
37
3012
.79
673.
09
4417
.11
1057
1160
21.2
514
6512
EG
AR
167.
844
298.
410
3012
.99
689.
410
6813
.55
736.
69
6815
.74
930.
910
EM
A5
323
.62
1713
1229
21.7
015
1112
3621
.22
1463
1211
320
.75
1414
11E
MA
100
129
.32
2349
1329
26.4
020
1713
3626
.59
2038
1359
24.7
318
3313
GA
RT
234.
395
99.0
04
745.
563
158.
96
696.
759
228.
87
116
10.1
346
0.5
7M
GA
R23
0.64
02.
346
837
00.
695
1.63
85
366
1.02
90.
013
511
92.
754
32.8
35
Copyright 2003 John Wiley & Sons, Ltd. J. Forecast. 22, 1–22 (2003)
20 C. Brooks and G. Persand
The first point to note is that if the objective is to cover 99% of future losses, then almost none ofthe models are adequate. The proportion of exceedences for the bond, share and commodity assetsis always considerably in excess of 1%—typically 1.4–2%. Thus for example, even the best modelat the 1-day horizon for the commodities data, which is the long term mean, has nearly 70% moreviolations of value at risk in the hold-out sample than would be expected under the null. Also forthis series, the majority of models have a TUFF statistic that takes on a value of one—that is, theyfail at the first observation! Almost none of the models for any of the four asset classes makes itto the hundredth observation, the expected time until first failure. Consequently, the UCF statisticrejects all models for all individual asset series at all horizons.
Matters are improved somewhat for the portfolio of assets, presumably as a result of the benefitsof diversification in reducing the number of extreme observations that lead to an exceedence ofthe VaR. The typical proportion of exceedences is reduced to around 1.2%, and although onlythe multivariate GARCH model has fewer than 1% exceedences, several models are acceptableaccording to the UCF test statistic. Similar patterns are revealed at the 1-day and longer horizons.The models fare much better when only 95% coverage is desired; more than half of the modelsachieve their nominal rate. In terms of model rankings, the long-term mean and the linear regressionin volatility models seem preferable, although again, there is no uniformly most accurate model.The GARCH model seems to provide reasonably accurate VaR estimates, evidenced by its actualcoverage rate being close to the nominal rate, although there is a tendency to over-estimate theVaR, a result also observed by Brooks, Clare and Persand (2000).
CONCLUSIONS
This paper has sought to re-examine the volatility forecasting literature in the context of a relativelynew use of volatility forecasts—for financial (market) risk assessment. A number of our resultsare worthy of further note. First, the gain from using a multivariate GARCH model for forecastingvolatility, which has not been previously investigated, is minimal. This result is true both understandard statistical and risk management evaluation measures. Given the complexity, estimationdifficulties, and computer-intensive nature of MGARCH modelling, we conjecture that unless theconditional covariances are required, the estimation of multivariate GARCH models is not worthwhile. In the context of portfolio volatility, more accurate results can be obtained by aggregating theportfolio constituents into a single series, and forecasting that, rather than modelling the individualcomponent volatilities and the correlations between the returns.
Second, it appears that some models are poor performers irrespective of both the series on whichthey are estimated and the loss function used to evaluate their forecasts. The random walk involatility, the EGARCH and to a lesser extent the EWMA models, fall into this category.
When it comes to selecting the ‘best’ model for forecasting, however, the particular evaluationmeasure employed plays a predominant role. Whilst there seems to be little difference in the modelrankings when the ex post measure is changed from low-frequency to high-frequency squaredreturns, the differences between rankings under statistical and risk management procedures aresubstantial. Although generalizing across data series (asset classes) and investment horizons isdifficult, overall the statistical measures preferred the GARCH(1,1) model over simpler techniquesand over its extensions and variants. On the other hand, when evaluated in the context of VaRestimates which achieve an appropriate out-of-sample coverage rate, the simplest models, such asthe long-term mean (historical average) or the autoregressive volatility model, are preferred. We
Copyright 2003 John Wiley & Sons, Ltd. J. Forecast. 22, 1–22 (2003)
Volatility Forecasting for Risk Management 21
thus concur with Dacco and Satchell (1999) in arguing that the choice of loss function can havean over-riding effect upon volatility forecasting accuracies; thus the debate on superior volatilityforecasting models should be considered far from resolved.
ACKNOWLEDGEMENTS
The authors would like to thank an anonymous referee for useful comments on a previous versionof this paper. The usual disclaimer applies.
REFERENCES
Akgiray V. 1989. Conditional heteroskedasticity in time series of stock returns: evidence and forecasts. Journalof Business 62: 55–80.
Alexander CO, Leigh CT. 1997. On the covariance models used in Value at Risk models. Journal ofDerivatives 4: 50–62.
Andersen T, Bollerslev T. 1998. Answering the skeptics: yes, standard volatility models do provide accurateforecasts. International Economic Review 39: 885–905.
Basle Committee on Banking Supervision. 1988. International Convergence of Capital Measurement andCapital Standards July.
Bollerslev T, Chou RY, Kroner KF. 1992. ARCH modelling in finance: a review of the theory and empiricalevidence. Journal of Econometrics 52: 5–59.
Bollerslev T, Engle RF, Wooldridge JM. 1988. A capital asset pricing model with time-varying covariances.Journal of Political Economy 96: 116–131.
Brailsford TJ, Faff RW. 1996. An evaluation of volatility forecasting techniques. Journal of Banking andFinance 20: 419–438.
Brooks C. 1998. Forecasting stock return volatility: does volume help? Journal of Forecasting17: 59–80.
Brooks C, Clare AD, Persand G. 2000. A word of caution on calculating market-based minimum capital riskrequirements. Journal of Banking and Finance 14(10): 1557–1574.
Brooks C, Henry OT, Persand G. 2002. Optimal hedging and the value of news. Journal of Business 75(2):333–352.
Brooks C, Persand G. 2000a. Value at Risk and market crashes. Journal of Risk 2(4): 5–26.Brooks C, Persand G. 2000b. Lies, damned lies and Value at Risk estimates. Risk 13(5): May, 63–66.Dacco R, Satchell S. 1999. Why do regime-switching models forecast so badly? Journal of Forecasting
18: 1–16.Day TE, Lewis CM. 1992. Stock market volatility and the information content of stock index options. Journal
of Econometrics 52: 267–287.Dimson E, Marsh P. 1990. Volatility forecasting without data snooping. Journal of Banking and Finace 14:
399–421.Dowd K. 1998. Beyond Value at Risk: The New Science of Risk Management. Wiley: Chichester.Embrechts P, Resnick SI, Samorodnitsky G. 1999. Extreme value theory as a risk management tool. North
American Actuarial Journal 3: 30–41.Fair RC, Shiller RJ. 1990. Comparing information in forecasts from econometric models. American Economic
Review 80: 375–389.Franses PH, van Dijk D. 1996. Forecasting stock market volatility using non-linear GARCH models. Journal
of Forecasting 15: 229–235.Heynen RC, Kat HM. 1994. Volatility prediction: a comparison of the stochastic volatility, GARCH(1,1) and
EGARCH(1,1) models. Journal of Derivatives 2: 50–65.Huisman R, Koedijk KG, Poqwnall RAJ. 1998. VaR-x: fat tails in financial risk management. Journal of Risk
1: 47–61.
Copyright 2003 John Wiley & Sons, Ltd. J. Forecast. 22, 1–22 (2003)
22 C. Brooks and G. Persand
Jackson P, Maude DJ, Perraudin W. 1998. Testing Value at Risk approaches to capital adequacy. Bank ofEngland Quarterly Bulletin 38: 256–266.
Johansen A, Sornette D. 1999. Critical crashes. Risk 12: 91–95.Jorion P. 1996. Value at Risk: The New Benchmark for Controlling Market Risk. Irwin: Chicago, IL.Morgan JP. 1996. Riskmetrics Technical Document 4th edn.Kupiec P. 1995. Techniques for verifying the accuracy of risk measurement models. Journal of Derivatives
2: 73–84.Pagan AR, Schwert GW. 1990. Alternative models for conditional stock volatilities. Journal of Econometrics
45: 267–290.West KD, Cho D. 1995. The predictive ability of several models of exchange rate volatility. Journal of
Econometrics 69: 367–391.West KD, Edison HJ, Cho D. 1993. A utility-based comparison of some models of exchange rate volatility.
Journal of International Economics 35: 23–45.
Authors’ biographies:Chris Brooks is Professor of Finance at the ISMA Centre, University of Reading, where he also obtained hisPhD. He has published over forty papers in leading academic and practitioner journals in finance, econometricsand economics, and is also author of the textbook Introductory Econometrics for Finance, published byCambridge University Press.
Gita Persand is a Lecturer in Finance at the Department of Economics, University of Bristol. She obtainedher PhD from the ISMA Centre, University of Reading. Her research interests are in the field of financial riskmanagement, and she has published in outlets such as Journal of Business, Journal of Banking and Finance,and Journal of Risk.
Authors’ addresses:Chris Brooks, ISMA Centre, University of Reading, PO Box 242, Whiteknights, Reading RG6 6BA, UK.
Gita Persand, Department of Economics, University of Bristol, Bristol B58, 1TH, UK.
Copyright 2003 John Wiley & Sons, Ltd. J. Forecast. 22, 1–22 (2003)