The Predictive Power of Value-at-Risk Models in Commodity Futures Markets

This paper examines the in- and out-of-sample performance of various value-at-risk (VaR) approaches for commodity futures investments: conventional VaR, the Cornish-Fisher (CF) VaR, GARCH-type VaR models, and semi-parametric conditional autoregressive value-at-risk (CAViaR) models, which do not depend on the assumption of normally distributed i.i.d. error terms. A model comparison reveals that determining the best VaR model depends strongly on the underlying return series. Our results suggest that the CAViaR and GARCH-type models generally outperform the other VaRs. These models can incorporate time-varying volatility adequately and are sensitive to changes in the return-generating process. This has important implications for the risk management of portfolios involving passive long-only commodity futures positions with heavy-tailed data-generating processes