Sampling and validation methods for hazard estimation

In this poster we compare in terms of predictive performance sampling and validation methods for hazard estimation in credit risk assessment. More precisely, we compare proportional hazard models based on the Cox model (Cox, 1972) with non parametric survival approaches based on random survival forest (Ishwaran et al., 2008) under different validation settings. On the basis of predictive performance measures (i.e. Breir Score and Prediction Error), we compare validation techniques based on the Bootstrap: Bootstrap without replacement (B=200), BOOT 632 (see e.g. Davison and Hinkley, 1997) and BOOT 632 PLUS (Efron and Tibshirani, 1997). Empirical evidence are given on a real data set provided by a credit rating agency composed of 742 small and medium enterprises, 9 financial ratios, a binary dependent variable which express a solvency indicator and a duration indicator