Forecasting Bankruptcy with Incomplete Information

We propose new specifications that explicitly account for information noise in the input data of bankruptcy hazard models. The specifications are motivated by a theory of modeling credit risk with incomplete information (Duffie and Lando [2001]). Based on over 2 million firm-months of data during 1979-2012, we demonstrate that our proposed specifications significantly improve both in-sample model fit and out-of-sample forecasting accuracy. The improvements in forecasting accuracy are persistent throughout the 10-year holdout periods. The improvements are also robust to empirical setup, and are more substantial in cases where information quality is a more serious problem. Our findings provide strong empirical support for using our proposed hazard specifications in credit risk research and industry applications. They also reconcile conflicting empirical results in the literature.