State-Dependent Structural Shocks and Temporal Generalization in Corporate Default Prediction

Abstract

This study examines whether financially structured data augmentation can enhance corporate default prediction under strict temporal generalization. Using an anchored temporal split—1999–2004 for training, 2005–2006 for validation, 2007–2009 for stress testing, and 2010–2018 for post-stress evaluation—it first documents a limitation of current practice: static augmentation fails to improve out-of-time discriminative power. The raw baseline attains a future mean AUC of 0.6858, whereas SMOTE and static con- strained augmentation yield only 0.6790 and 0.6798. The paper then introduces a state- dependent structural shock framework that injects economically disciplined deterioration paths into stable-regime training data. The calibrated acute/persistent design achieves the highest future mean AUC, 0.6947, above the regime-feature baseline of 0.6927. Boot- strap evidence supports this AUC gain directionally rather than uniformly, with stronger support in the post-stress window than in the stress window. By contrast, the gain does not extend to future mean PR-AUC. Decision simulation further shows that the practical advantage of shocks is strongest under conservative screening rules rather than aggressive recall. Dynamic stress rehearsal therefore improves out-of-time ranking selectively, but not universally across metrics, baselines, or operating points.