Forecasting liquidity-adjusted intraday Value-at-Risk with vine copulas

Download Forecasting liquidity-adjusted intraday Value-at-Risk with vine copulas

Post on 20-Dec-2016

219 views

Category:

Documents

3 download

Embed Size (px)

TRANSCRIPT

<ul><li><p>al</p><p>Keywords:</p><p>intl DoBy, waskprebe</p><p> 2013 Elsevier B.V. All rights reserved.</p><p>to botharket ts to exiassetsBrennaosed t</p><p>based on vine copulas for estimating and forecasting liquidity-adjusted risk measures for multivariate stock portfolios and applythe proposed model on returns and bid-ask spreads reconstructedfrom high frequency data. We start our analysis by performing avariety of diagnostic tests on the dependence structure betweenintraday stock returns and quoted bid-ask spreads for selected</p><p>el for returseveral ty</p><p>liquidity-adjusted risk measures to illustrate the superiorityproposed method.</p><p> Corresponding author. Tel.: +49 231 755 4608.E-mail addresses: gregor.weiss@tu-dortmund.de (G.N.F. Wei), hendrik.supper@</p><p>tu-dortmund.de (H. Supper).</p><p>2 Obviously, other measures of liquidity could have also been used. The majority ofstudies in the risk management literature, however, regularly employs the quotedbid-ask spread to measure liquidity (see e.g. Berkowitz, 2000; Bangia et al., 2002).</p><p>3 The relationship between the returns and the bid-ask spread of the samecompany, on the other hand, is well documented, see e.g. Amihud and Mendelson(1986) who nd that average portfolio risk-adjusted returns increase with their bid-ask spread with the slope of the return-spread relationship decreasing with thespread.</p><p>Journal of Banking &amp; Finance 37 (2013) 33343350</p><p>Contents lists available at SciVerse ScienceDirect</p><p>Journal of Banking &amp; Finance</p><p>w.elsevier .com/locate / jbf1 Tel.: +49 231 755 8212.et al., 2012).In this paper, we propose a multivariate econometric model</p><p>stocks. Based on our estimated multivariate modbid-ask spreads, we then forecast and backtest0378-4266/$ - see front matter 2013 Elsevier B.V. All rights reserved.http://dx.doi.org/10.1016/j.jbankn.2013.05.013ultiplens andpes ofof ourin liquidity (Acharya and Pedersen, 2005; Sadka, 2006). Whilethe nance literature has long been concerned with liquidity riskand market microstructure (Demsetz, 1968; Stoll, 1978; Amihudand Mendelson, 1980), the recent nancial crisis has renewedinterest in both the modeling of liquidity risk and in the analysisof its driving factors (see e.g. Cornett et al., 2011; Dick-Nielsen</p><p>ask spreads stemming from different companies has not been ana-lyzed before in the literature,3 our dependence model is requiredto be able to capture a wide range of possible linear and non-lineardependencies. We therefore resort to the concept of vine copulas(Heinen and Valdesogo, 2008; Aas et al., 2009) to model the depen-dence structure between the returns and bid-ask spreads of mLiquidityCommonalityVine copulasLiquidity-adjusted intraday Value-at-Risk</p><p>1. Introduction</p><p>Liquidity risk is of major concernmanagers. Especially in times of mliquidity can dry up forcing investorcost. Consequently, investors prefer(see Amihud and Mendelson, 1986;1996; Liu, 2006) or are least not expinvestors and portfoliourmoil, overall markett positions at increasedwhich are either liquidn and Subrahmanyam,o systematic decreases</p><p>companies on the NASDAQ 100 in 2009.2 These preliminary testsprovide us with ample empirical evidence of not only strong linearcorrelation but also strong tail dependence between bid-ask spreadsacross individual stocks. We then turn to modeling the marginalbehavior of stock returns and bid-ask spreads by the use of GARCHprocesses and the Autoregressive Conditional Double Poisson model,respectively (Heinen, 2003; Gro-Klumann and Hautsch, 2011). Asthe dependence structure between intraday stock returns and bid-G12G14</p><p>motivating our use of a vine copula model. Furthermore, the backtesting results for the L-IVaR of a port-folio consisting of ve stocks listed on the NASDAQ show that the proposed models perform well in fore-casting liquidity-adjusted intraday portfolio prots and losses.Forecasting liquidity-adjusted intraday V</p><p>Gregor N.F. Wei , Hendrik Supper 1Technische Universitt Dortmund, Otto-Hahn-Str. 6a, D-44227 Dortmund, Germany</p><p>a r t i c l e i n f o</p><p>Article history:Received 29 June 2012Accepted 4 May 2013Available online 14 May 2013</p><p>JEL classication:C58C53</p><p>a b s t r a c t</p><p>We propose to model the joAutoregressive Conditionathe dependence structure.spreads from intraday dataements of stocks and bid-Value-at-Risk (L-IVaR). In aand strong tail dependence</p><p>journal homepage: wwue-at-Risk with vine copulas</p><p>distribution of bid-ask spreads and log returns of a stock portfolio by usinguble Poisson and GARCH processes for the marginals and vine copulas forestimating the joint multivariate distribution of both returns and bid-aske incorporate the measurement of commonalities in liquidity and comov-spreads into the forecasting of three types of liquidity-adjusted intradayliminary analysis, we document strong extreme comovements in liquiditytween bid-ask spreads and log returns across the rms in our sample thus</p></li><li><p>Although early empirical and theoretical studies on liquidityhave concentrated on individual securities, the analysis of comov-ements in liquidity among individual stocks has become a cornerstone of the microstructure literature starting with the seminal pa-</p><p>G.N.F. Wei, H. Supper / Journal of Bankinpers by Chordia et al. (2000), Hasbrouck and Seppi (2001) andHuberman and Halka (2001). Since then, studies on the common-ality in liquidity have unanimously found clear empirical evidencefor strong comovements in liquidity as proxied, e.g., by the bid-askspreads of individual stocks. The determinants driving thiscommonality in liquidity, however, have remained relatively un-known until the recent study by Karolyi et al. (2012). They nd thatcommonality in liquidity is positively correlated with high marketvolatility. In addition, liquidity commonalities were particularlystrong during the Asian crisis and the recent nancial crisisunderlining the necessity of taking liquidity risk into account inmultivariate risk modeling.</p><p>The incorporation of liquidity risk into the measurement ofmarket risk has been a recurring topic in the risk management lit-erature since the onset of the Value-at-Risk (VaR) concept as a defacto industry standard. Although standard VaR lacks a rigorousconsideration of liquidity risk, several extensions to certain formsof Liquidity-Adjusted VaRs (L-VaR) have been proposed in the liter-ature (see e.g. Berkowitz, 2000; Bangia et al., 2002; Qi and Ng,2009). Parallel to the study of liquidity-adjusted risk measures,the availability of tick-by-tick data has enabled researchers tomeasure the liquidity of stocks at an ultra-high frequency (Chordiaet al., 2000; Engle, 2000). While the increase in the speed of tradingand the widespread availability of transaction data have led to anincrease in the importance of intraday risk analysis (Dionneet al., 2009; Gourieroux and Jasiak, 2010), the modeling of intradaystock returns and bid-ask spreads is severely hampered by theirregular spacing of such data. Econometric research has thus con-centrated on deriving appropriate models for the duration betweentransactions (Engle and Russell, 1998), for the estimation of intra-day volatility (Andersen et al., 2000, 2003) and for the estimationof intraday risk measures (Giot, 2005; Dionne et al., 2009).</p><p>The econometrics literature includes several studies whichfocus on modeling and forecasting multivariate intraday stockreturns or multivariate bid-ask spreads (see e.g. Engle and Russell,1998; Breymann et al., 2003; Heinen and Rengifo, 2007;Gro-Klumann and Hautsch, 2011; Li and Poon, 2011). Up to date,however, no study has analyzed the joint distribution of intradayreturns and bid-ask spreads of multiple stocks. Although stock re-turns are known to be leptokurtic and non-normally distributedwith time-varying dependence (see e.g. Longin and Solnik, 1995),and bid-ask spreads are known to comove across markets, littleis known about the dependence (and consequently possible com-ovements) between liquidity and returns of different stocks.</p><p>In addition, the multivariate modeling of bid-ask spreads has sofar been a purely econometric exercise with no connection to therich literature on liquidity commonality.4 Furthermore, there existonly few studies on the estimation of VaR from intraday data(Dionne et al., 2009) and, to the best knowledge of the authors, nowork on multivariate liquidity-adjusted portfolio-VaR. Our papertries to link the research on liquidity commonality and on theestimation of multivariate portfolio-VaR from high frequency data.</p><p>The contributions of this study relative to the existing literatureon liquidity commonality and liquidity risk management aresignicant and numerous. While previous studies on the</p><p>4 As stated by Heinen and Rengifo (2007), extending the standard AutogressiveConditional Duration (ACD) model of Engle and Russell (1998) to more than one timeseries has proven to be quite difcult. Considering the numerous problems one</p><p>encounters when trying to model multivariate count data, it is, however, notsurprising that studies such as the one by Heinen and Rengifo (2007) focus on theeconometric side of the problem.Loosely speaking, liquidity can be seen as the ability of a marketto allow for immediate trading even large amounts tominimal costswithout causing remarkable price movements; hence, liquidity riskresults from the difference between transaction and market price.</p><p>Kyle (1985) provides a formal denition and introduces threecomponents of liquidity including tightness, depth and resiliency;according to Amihud and Mendelson (1986), a natural measure ofliquidity (or rather liquidity risk) is the spread between the bid andask prices. Bangia et al. (2002) classify liquidity risk into exogenousliquidity risk which arises from the general conditions of a marketand is equal to all participants and endogenous liquidity risk whichrefers to the volume of individual trading positions and which isidiosyncratic.</p><p>The bid-ask spread has become a key parameter in modelingnancial data and captures the costs of immediate trading, whichcan be explained by the liquidity suppliers purchasing at the bidand selling at a higher ask price in order to recoup their own costs.Empirical properties like intraday seasonalities, linear dependenceand commonalities are given in the context of the data descriptionin Section 3.</p><p>Statistically, bid-ask spreads belong to the class of discretecount data, because they count the number of ticks between bidand ask prices. To accurately model time series of count data mea-sured at high frequency, there are two important aspects toconsider: on the one hand the model has to cope with the issuesThe purpose of this section is to outline the econometric modelsfor intraday bid-ask spreads and returns, their multivariate depen-dence structure as well as liquidity-adjusted intraday VaR.</p><p>We start with the models for intraday spreads and returnswhich form the fundamental building blocks of the VaR modelsdescribed later.</p><p>2.1. The autoregressive conditional double poisson modelcommonality in liquidity have documented strong positive linearcorrelation between measures of liquidity, this study is the rstto nd empirical evidence for a strong non-linear dependence inthe form of signicant tail dependence as well. This paper is alsothe rst to integrate our understanding of liquidity commonalityinto the estimation and forecasting of liquidity-adjusted risk mea-sures. In addition, this paper constitutes the rst multivariate mod-el for the joint distribution of the returns and bid-ask spreads ofmultiple stocks. Finally, our paper adds to the fastly growing liter-ature on the use of vine copulas in risk management and asset pric-ing applications making full use of the vines exibility.</p><p>The results presented in this paper show that the proposedmul-tivariate model for bid-ask spreads and intraday returns performsexceptionally well in forecasting liquidity-adjusted portfolio losses.While losses are adequately bounded below by our liquidity-ad-justed VaR forecasts, our models are not too conservative and thusprevent investors like, e.g., banks from reserving unnecessary riskcapital buffers. At the same time, we show that the neglection ofliquidity risk and non-linearities in the dependence between re-turns and bid-ask spreads in the forecasting of portfolio-VaRs canlead to a severe underestimation of losses on the portfolio.</p><p>The remainder of this article is structured as follows. Section 2presents the econometric methodology. Section 3 discusses theempirical study as well as the results. Concluding remarks aregiven in Section 4.</p><p>2. Econometric methodology</p><p>g &amp; Finance 37 (2013) 33343350 3335of discreteness and serial dependence, on the other hand itsestimation procedure has to be tractable for a large number ofobservations. Following Gro-Klumann and Hautsch (2011), we</p></li><li><p>drop the traditional approaches reviewed by Cameron and Trivedi(1998), MacDonald and Zucchini (1997) and McKenzie (2003)because of their complex estimation procedures and adopt theAutoregressive Conditional Double Poisson Model (ACDP hereaf-ter) that belongs to the class of observation-driven models and isbased on the Autoregressive Conditional Poisson Model (ACP)</p><p>3336 G.N.F. Wei, H. Supper / Journal of Banking &amp; Finance 37 (2013) 33343350introduced by Rydberg and Shephard (1999).5</p><p>Since the Poisson distribution is a natural starting point forcounts (see Heinen, 2003), the ACP assumes that the counts followa Poisson distribution with an autoregressive mean. Extending thisapproach, the ACDP uses the Double Poisson distribution as pro-posed by Efron (1986) instead and allows for a more exible mod-eling of the spreads.6 To provide a mathematical description of thismodel we start with some notations: letM;N;d 2 N denote the num-ber of trading days in the sample, the number of observations perday and the number of stocks, respectively, and let I, J, K be indexsets given by I : 1;M \N; J : 0;N 1 \N0 and K : 1;d \N,where N0 : N [ f0g. Further, let N0;N1 2 N [ f0g with N0 &lt; N1 andlet Z : fN0 n0 &lt; n1 &lt; . . . &lt; nN1 6 N1g be a partition of [N0,N1]with equidistant points nj 2 N0j 2 J in ascending order and normDZ , denoting the time grid of observation points for the sample</p><p>days.7 Let Aki;j and Bki;j denote the (best) bid and the (best) ask price</p><p>(respectively) and Ski;j : Aki;j Bki;j the quoted bid-ask spread of stockk closest to time nj 2 Z, realized at the ith trading day (i 2 I,k 2 K).Moreover, let DPois (k,c) be the Double Poisson distribution (Efron,1986) and F ki;j the information available on the series of stock k upto and including time nj of the ith day.8 Let L denote the lag operator</p><p>dened by LbSki;j : Ski;jb for b 2 Z and let two lag polynomials Uk andWk be given as UkL :</p><p>Pqm1/</p><p>kmL</p><p>m and WkL :Pp</p><p>n1wknL</p><p>n, wherep; q 2 N0 refer to the order of the process and the coefcients arerestricted by /m &gt; 0, m = 1, . . . , q as well as w</p><p>kn &gt; 0, n = 1, . . . , p.</p><p>Then, the ACDP (p,q) process Ski;jn o</p><p>j2Jis dened by</p><p>Ski;jjF ki;j1 DPois kki;j; ck </p><p>;</p><p>kki;j ck UkLSki;j WkLkki;j;1</p><p>where i 2 I, k 2 K and ck &gt; 0.9The ACDP assumes the...</p></li></ul>