Stress Testing Credit Risk Models: Algorithmics Mark-to-Future
A key issue for bankers and regulators is internal model validation and predictive accuracy. In the context of market models, this issue has led to numerous efforts to back-test models to ascertain their predictive accuracy. The second pillar of the Basel II capital accords states that bank regulators must evaluate how well banks are assessing their capital needs relative to their risk, thereby requiring bank examiners to validate the accuracy of bank risk measurement models. Currently, under the Basel market risk-based capital requirements, a bank must back-test its internal market model over a minimum of 250 past days if it is used for capital requirement calculations. If the forecast VAR errors on those 250 days are too large, implying that the risk is underestimated on too many days, a system of penalties is imposed by regulators to create incentives for bankers to get their models right.1
Many observers, however, have argued that back-testing over 250 days is simply not enough, given the high standard errors that are likely to occur if the period is not representative of true market conditions. To reduce errors of this type, one suggestion has been to increase the number of past daily observations over which a back-test of a model is conducted. For example, using at least 1,000 past daily observations is commonly considered to be adequate to ensure that the period chosen is representative ...