We evaluate predictive performance of a selection of value-at-risk (VaR) models for Japanese stock market data. We consider traditional VaR models such as Riskmetrics method, historical simulation, variance-covariance method, Monte Carlo method, and their variants which are integrated with various ARCH models. Also considered are more recent models based on non-parametric quantile regression and extreme value theory (EVT). We apply these methods to the Japanese stock market index (1984-2000) and compare their performances in terms of various evaluation criteria using the method of White [Econometrica 68 (5) (2000) 1097-1126] for three out-of-sample periods of 19951996, 1997-1998, and 1999-2000. (C) 2002 Elsevier Science B.V. All rights reserved.