Value-at-risk and the global financial crisis

Using daily data for seven large international banks, we examine the forecasting ability of bank value-at-risk (VaR) estimates around the 2007-9 global financial crisis (GFC) period. We find that the banks' internal VaR estimates are very inaccurate. They systematically overstated VaR during the pre- and postcrisis periods, with mixed performance during the GFC. Some banks inflated their VaRs, while others experienced excessive VaR exceptions and clustering. VaR estimates based on simple models of generalized autoregressive conditional heteroscedasticity (GARCH) type easily outperform internal VaR estimates. The VaR estimated via a GARCH-t distribution captures the extreme losses reasonably well. We attribute the poor VaR estimates at banks to the banks' inappropriate choice of internal VaR models.