Forecasting Var Using Analytic Higher Moments for GARCH Processes
Total Page:16
File Type:pdf, Size:1020Kb
Forecasting VaR using analytic higher moments for GARCH processes Article (Unspecified) Alexander, Carol, Lazar, Emese and Silvia, Stanescu (2013) Forecasting VaR using analytic higher moments for GARCH processes. International Review of Financial Analysis, 30. pp. 36-45. ISSN 1057-5219 This version is available from Sussex Research Online: http://sro.sussex.ac.uk/id/eprint/45651/ This document is made available in accordance with publisher policies and may differ from the published version or from the version of record. If you wish to cite this item you are advised to consult the publisher’s version. Please see the URL above for details on accessing the published version. Copyright and reuse: Sussex Research Online is a digital repository of the research output of the University. Copyright and all moral rights to the version of the paper presented here belong to the individual author(s) and/or other copyright owners. To the extent reasonable and practicable, the material made available in SRO has been checked for eligibility before being made available. Copies of full text items generally can be reproduced, displayed or performed and given to third parties in any format or medium for personal research or study, educational, or not-for-profit purposes without prior permission or charge, provided that the authors, title and full bibliographic details are credited, a hyperlink and/or URL is given for the original metadata page and the content is not changed in any way. http://sro.sussex.ac.uk Author's personal copy International Review of Financial Analysis 30 (2013) 36–45 Contents lists available at SciVerse ScienceDirect International Review of Financial Analysis Forecasting VaR using analytic higher moments for GARCH processes Carol Alexander a,⁎, Emese Lazar b, Silvia Stanescu c a School of Business, Management and Economics, University of Sussex, Falmer, Brighton, Sussex BN1 9SL, UK b ICMA Centre, University of Reading, Reading RG6 6BA, UK c Kent Business School, University of Kent, Canterbury CT2 7PE, UK article info abstract Article history: It is widely accepted that some of the most accurate Value-at-Risk (VaR) estimates are based on an appropri- Received 10 June 2012 ately specified GARCH process. But when the forecast horizon is greater than the frequency of the GARCH Received in revised form 25 April 2013 model, such predictions have typically required time-consuming simulations of the aggregated returns distri- Accepted 28 May 2013 butions. This paper shows that fast, quasi-analytic GARCH VaR calculations can be based on new formulae for Available online 6 June 2013 the first four moments of aggregated GARCH returns. Our extensive empirical study compares the Cornish– Fisher expansion with the Johnson SU distribution for fitting distributions to analytic moments of normal JEL classification: fi C53 and Student t, symmetric and asymmetric (GJR) GARCH processes to returns data on different nancial as- G17 sets, for the purpose of deriving accurate GARCH VaR forecasts over multiple horizons and significance levels. © 2013 Elsevier Inc. All rights reserved. Keywords: GARCH Higher conditional moments Approximate predictive distributions Value-at-Risk S&P 500 Treasury bill rate Euro–US dollar exchange rate 1. Introduction (“top–down”) level, rather than utilizing standard (“bottom–up”)VaR models for assessing a firm's market risk capital. A path-breaking Since the 1996 Amendment to the Basel I Accord, Value-at-Risk paper by Berkowitz and O'Brien (2002) utilizes aggregate profitand (VaR) has become the standard metric for financial risk assessment loss data from six of the world's major banks to demonstrate a very and reporting, not only in the major banks that must now use VaR clearly superior accuracy in top–down GARCH-based VaR estimates rel- forecasts as a basis for their assessment of market risk capital re- ative to more traditional, bottom–up VaR estimates.2 serves, but also in asset management, hedge funds, mutual funds, An α% n-day VaR estimate is the loss that will not be exceeded, pension funds, corporate treasury and indeed in virtually every with a (1 − α)% level of confidence, if the portfolio is left unmanaged large institution worldwide that has dealings in the financial markets. over a period of n days. When VaR is quoted as a percentage the cur- As a result the academic literature on forecasting VaR is huge.1 rent portfolio value, it may therefore be derived from the α-quantile Given the widely documented characteristics of financial asset of the n-period portfolio return distribution, as: returns, quite complex dynamic models are needed for predicting − their distributions. A salient feature is their volatility clustering and gen- VaRn;α;t ¼ −^−1 ðÞα ; ^ ðÞ ¼ α ð Þ eralised autoregressive conditional heteroscedastic (GARCH) models, VaRn;α;t F t;tþn or equivalently as ∫ f t;tþn x dx 1 introduced by Engle (1982), Bollerslev (1986) and Taylor (1986), −∞ have proved very successful in capturing this behaviour. Such models ^−1 can also partially explain why asset returns distributions are skewed where F t;tþn is the time t forecast of the inverse distribution function and leptokurtic. Some of the most influential academic research con- (also called quantile function) for the returns aggregated from time t cerns the use of GARCH processes to forecast VaR at the aggregate 2 Perignon and Smith (2010) note that historical simulation is the most widely-used approach, based on a survey of major banks around the world. However, Alexander ⁎ Corresponding author. Tel.: +44 1273 873950; fax: +44 1273 873715. and Sheedy (2008) demonstrate that historical simulation is highly inaccurate without E-mail addresses: [email protected] (C. Alexander), additional ‘filtering’—see, for example, Barone-Adesi, Bourgoin, & Giannopoulos, [email protected] (E. Lazar), [email protected] (S. Stanescu). (1998), and Barone-Adesi, Giannopoulos, & Vosper, (1999), where the historical VaR 1 Recent reviews may be found in Alexander (2008), Angelidis and Degiannakis methodology is augmented with a GARCH model. See also Boudoukh, Richardson, (2009) and Christoffersen (2009). and Whitelaw (1998), for an alternative filtering approach. 1057-5219/$ – see front matter © 2013 Elsevier Inc. All rights reserved. http://dx.doi.org/10.1016/j.irfa.2013.05.006 Author's personal copy C. Alexander et al. / International Review of Financial Analysis 30 (2013) 36–45 37 ^ to time t + n,andf t;tþn is the corresponding density function. For the model by Alexander, Lazar, and Stanescu (2011), formulae which are purpose of VaR estimation the GARCH model is usually estimated using similar but not identical to those derived by Wong and So (2003).We daily data. However, for many empirical applications – and especially prefer to use moments of the asymmetric GJR-GARCH(1,1) process for computing regulatory capital to cover market risks in banks, which with a generic conditional distribution, instead of the AGARCH (p, q) is typically based on a 10-day VaR estimate derived from daily data – model considered by Wong and So (2003), because the former model we are often interested in longer horizons. The problem is that for encompasses the majority of the GARCH models that are favoured in ^ ^−1 fi n > 1 neither F t;tþn nor F t;tþn is known in closed form (in a GARCH con- the nancial forecasting literature: see Awartani and Corradi (2005), text based on daily data) so they are obtained using simulations. This is in Asai and McAleer (2008) and many others. accordance with Engle (2003), who argued in his Nobel lecture that sim- The reminder of this paper is organised as follows: Section 2 pre- ulations are required to predict the quantiles of the returns distribution sents the theoretical methodology that we shall implement for our over a time horizon which is longer than the frequency of the model, empirical results and explains how analytic formulae for the first when aggregated returns are generated by a GARCH process. But simula- four moments of aggregated GARCH returns can be used to approxi- tions are only asymptotically exact, so it can be very time consuming to mate VaR; Section 3 presents the data and empirical results; and simulate aggregated GARCH returns distributions that allow VaR to be Section 4 concludes. forecast with a satisfactory degree of accuracy. This computational burden has been an impediment to the adop- 2. Analytic approximations for GARCH VaR tion of VaR models based on GARCH processes in practice. Further- more, from an academic perspective, it has reduced the scope for We construct quasi-analytic VaR estimates that capture the im- extensive out-of-sample tests of GARCH-based VaR forecasts. portant characteristics of financial asset returns (i.e. their volatility Hence, the need arises for an alternative method of resolution that is clustering and non-normal distributions) by applying established less time-consuming than simulation, while retaining the great advan- moment-based approximation methods to analytic formulae for the tage of accurate GARCH modelling. Given the frequent turmoil in finan- first four conditional moments of GARCH aggregated returns. cial markets and the pervasive use of the VaR metric throughout the Consider the following generic GJR specification, introduced by industry, the construction of fast, accurate and easily implemented Glosten et al. (1993), for the generating process of a continuously VaR forecasts is of significant practical and regulatory importance. compounded portfolio return from time t − 1 to time t, denoted rt: In this paper we forecast aggregated returns distributions using ana- lytic formulae for the higher-order conditional moments of GARCH aggre- ¼ μ þ ε ; ε ¼ 1=2; ∼ ðÞ; ; rt t t zt ht zt D 0 1 gated returns. Given these moments we compare two VaR forecasts obtained using two different methods to approximate the future returns with3 distribution.