Why Don't Stress Tests Break Banks?

Abstract

Despite continued advances in the state of the art, stress tests predict failure less often than would seem reasonable. This paper argues that even if a stress test produces a good estimate for aggregate losses, failing to account for unevenness of losses could easily lead to an under-estimate of likely failures. Instead of relying on mean estimates for residential loan portfolio default rate, LVR and underlying property values, this paper uses distributional properties, providing a more realistic setting by allowing for variation between banks. A key finding is that once such variation is included, the increase in average banking system losses is substantial. Both the average bank and the worst performing group of banks suffer a significantly greater magnitude of loss. A simple intensification algorithm is then presented which allows for stress tests to generate more insight into the likelihood of failure for a given population under stress. The intensification outcomes may well generate a better specified stress model, as the failure patterns brought to light by intensification may reveal better ways to differentiate the stressed population.