CONDITIONAL VALUE AT RISK.pdfVIP

CONDITIONAL VALUE AT RISKDirk OrmoneitDepartment of StatisticsStanford UniversityStanford, CA 94305-4065ormoneit@ Ralph NeuneierSiemens AG, ZT IK 481730 MunchenGermanyRalph.Neuneier@mchp.siemens.deAbstractWe suggest a new methodology to overcome several well-known decienciesof Value at Risk computations. Our approach mainly addresses two aspects ofValue at Risk: rst, to avoid potentially disastrous clustering in predicted tailevents we derive a new approach to accurately estimating the conditional distri-bution of asset returns using maximum entropy densities. Second, by the verynature of the maximum entropy model, we account for negative skewness andfat tails in asset returns. In particular, to obtain a robust and scalable estimateof the covariance matrix of the assets in the portfolio we extend an approachby Hull and White to the case of conditional distributions. In extensive exper-iments with historical stock index data we compare the proposed methodologyto alternative estimation approaches using a new, simulation-based statisticaltesting procedure for serial dependence in the predicted tail events.1 IntroductionAs value at risk (\VaR) has emerged as the dominant measure of market riskimplied by a portfolio over the past years, extensive research both within academiaand within the nancial industry has recently focussed on eective and reliable algo-rithms to compute the VaR of complicated portfolios. Probably the most in uentialcontribution in this eld has been J.P. Morgans RiskMetricsTM methodology, withinwhich a multivariate normal distribution is employed to model the joint distributionof the assets in a portfolio [Mor96]. While this approach is doubtlessly appealingfor its straightforwardness and computational simplicity, several authors have foundthat the observed frequency of extreme market movements is frequently higher thanpredicted by RiskMetrics (see, for example, [HW98], Exhibit 3). Obvious explana-tions for this nding are negative skewness and