Validation of Default Probabilities

Well-performing default predictions show good discrimination and calibration. Discrimination is the ability to separate defaulters from nondefaulters. Calibration is the ability to make unbiased forecasts. I derive novel discrimination and calibration statistics to verify forecasts expressed in terms of probability under dependent observations. The test statistics’ asymptotic distributions can be derived in analytic form. Not accounting for cross correlation can result in the rejection of actually well-performing predictions, as shown in an empirical application. I demonstrate that forecasting errors must be serially uncorrelated. As a consequence, my multiperiod tests are statistically consistent.