A1248
Title: Corporate default prediction under distribution shift
Authors: Suguru Yamanaka - Aoyama Gakuin University (Japan) [presenting]
Abstract: Distribution shifts, or covariate shifts, arising from changing economic environments pose a significant challenge to the estimation of corporate default prediction models. A default prediction model based on an importance-weighted least-squares probabilistic classifier (IWLSPC) is introduced to address this degradation in model performance. In contrast to the standard least-squares probabilistic classifier (LSPC), IWLSPC adapts to these shifts by incorporating data density ratios as importance weights in model estimation. Empirical validation using real corporate default datasets demonstrates that the default prediction model based on IWLSPC achieves significantly higher predictive accuracy than the conventional, unweighted LSPC. The findings confirm that importance weighting is a robust technique for building reliable credit risk models capable of operating under non-stationary conditions.
