Attention-driven environment-adaptive financial risk prediction for multi-period forecasting
TL;DR
Evaluation against baselines including LSTM, XGBoost, and Altman's Z-score using AUC and interpretability scores reveals that DEARPM achieves an AUC of 0.89 and 0.93 in recession and expansion periods, respectively—significantly outperforming benchmarks.
Abstract
This study proposes the Dynamic Environment-Adaptive Risk Prediction Model, an attention-based neural network architecture that integrates an environmental perception module with a dynamic indicator generation mechanism to address multi-period financial risk forecasting. Comparative experiments were conducted using multi-source datasets comprising China’s A-share listed companies’ financial reports and macroeconomic stress indices. Evaluation against baselines including LSTM, XGBoost, and Altman's Z-score using AUC and interpretability scores reveals that DEARPM achieves an AUC of 0.89 and 0.93 in recession and expansion periods, respectively—significantly outperforming benchmarks. Dynamic indicators significantly enhanced class separation, increasing the Mahalanobis distance between high-risk and standard class centroids by a factor of 3.2 and substantially improving the Fisher discriminant ratio compared to static features. Combined with SHAP-based feature attribution, this clear separation provides robust and objective interpretability for risk management under dynamic market conditions. The model demonstrates strong robustness by maintaining an AUC> 0.88 under 50% missing data and 25% noise, offering high practical value for risk management under dynamic market conditions.