Evaluating Value-at-Risk models with desk-level data

We present new evidence on disaggregated profit and loss and VaR forecasts obtained from a large international commercial bank. Our dataset includes daily P/L generated by four separate business lines within the bank. All four business lines are involved in securities trading and each is observed daily for a period of at least two years. Given this rich dataset, we provide an integrated, unifying framework for assessing the accuracy of VaR forecasts. A thorough Monte Carlo comparison of the various methods is conducted to provide guidance as to which of these many tests have the best finite-sample size and power properties. The Caviar test of Engle and Manganelli (2004) performs best overall but duration-based tests also perform well in many cases.