Dynamic polynomial gradient boosting for interpretable financial distress early warning

Liu, Wanan, Pang, Congyuan, Chen, Naixi, Zou, Yao, Lan, Xingyu, Ma, Hongqiang
Risk Management · Springer Nature · 2026-11-04 · DOI: 10.1057/s41283-026-00251-6

Аннотация

Identifying companies at risk of financial distress is challenging due to severe class imbalance and limited model interpretability. This study develops PolyBoost, an early-warning framework integrating a polynomial loss into extreme gradient boosting and coupling the model with TreeSHAP. The formulation enhances minority-class sensitivity on boundary-adjacent difficult instances without artificial resampling heuristics. Using Chinese listed firms alongside an independent Polish bankruptcy dataset, we evaluate PolyBoost across lead times up to four years in advance. For Chinese listed firms, PolyBoost achieves a geometric mean of 0.923 and detects 93.6% of distressed companies one year prior to distress, compared with 50.8% for standard extreme gradient boosting, while sustaining competitive discrimination over longer horizons. External tests on Polish firms confirm cross-market robustness with geometric means of 0.904 one year ahead and 0.848 four years ahead. PolyBoost supports earlier and more transparent financial distress supervision.

Читать у издателя →

Похожие статьи