Frequency-aware gradient modulated boosted trees for interpretable financial distress prediction

Corporate financial distress represents a critical instability phenomenon within the complex, nonlinear dynamical system of financial markets. Accurately predicting these rare phase transitions is essential for maintaining systemic economic stability, yet conventional methods often fail to capture the intricate nonlinear fluctuations and are severely constrained by the inherent sparsity of distress signals. To address these challenges, this paper proposes XGBoost-FAGM, an adaptive boosting ensemble framework designed to forecast financial distress under severe class imbalance. Unlike traditional cost-sensitive approaches that rely on static, exogenous constraints, we introduce a Frequency-Aware Gradient Modulation (FAGM) mechanism. This mechanism endogenously modulates the optimization landscape by perceiving the frequency of rare distress events, thereby dynamically adjusting prediction margins to capture critical signals in a non-equilibrium environment. Furthermore, we embed TreeSHAP to disentangle the complex, nonlinear interactions among risk factors, transforming black-box predictions into interpretable risk attractors. Empirical analysis on A-share listed companies (2000–2024) demonstrates that XGBoost-FAGM significantly outperforms existing ensemble paradigms in identifying distressed states, particularly in Gmean and True Positive Rate metrics. By integrating adaptive gradient modulation with nonlinear interpretability, this study provides a robust methodological framework for monitoring the stability of complex financial systems and identifying early warning signals of corporate failure.