Semi-nonparametric VaR forecasts for hedge funds during the recent crisis

• Traditional VaR measures fail to capture risk in highly volatile scenarios.
• Hedge funds demand accurate techniques for risk management.
• Semi-nonparametric methods and EVT capture risk accurately when volatility is high.
• Gram–Charlier copula incorporates non-linear dependences in portfolio risk assessment.
• Two-step estimation of short Gram–Charlier series simplifies the VaR computation.

The need to provide accurate value-at-risk (VaR) forecasting measures has triggered an important literature in econophysics. Although these accurate VaR models and methodologies are particularly demanded for hedge fund managers, there exist few articles specifically devoted to implement new techniques in hedge fund returns VaR forecasting. This article advances in these issues by comparing the performance of risk measures based on parametric distributions (the normal, Student’s tt and skewed-tt), semi-nonparametric (SNP) methodologies based on Gram–Charlier (GC) series and the extreme value theory (EVT) approach. Our results show that normal-, Student’s tt- and Skewed tt- based methodologies fail to forecast hedge fund VaR, whilst SNP and EVT approaches accurately success on it. We extend these results to the multivariate framework by providing an explicit formula for the GC copula and its density that encompasses the Gaussian copula and accounts for non-linear dependences. We show that the VaR obtained by the meta GC accurately captures portfolio risk and outperforms regulatory VaR estimates obtained through the meta Gaussian and Student’s tt distributions.