Explainable Financial Risk Prediction Models—A Hybrid Framework of XGBoost+Transformer Based on SHAP and Time Attention Mechanisms

Existing financial risk prediction research mostly employs static models or pure time-series models, which struggle to simultaneously achieve efficient utilization of structured features and precise capture of dynamic temporal patterns. Furthermore, interpretability is often limited to feature importance, lacking attribution capabilities across the time dimension. To improve the accuracy and decision support value of financial risk identification, this paper investigates a hybrid framework combining XGBoost (eXtreme Gradient Boosting) and Transformer, based on SHAP (SHapley Additive exPlanations) and a time attention mechanism. A Transformer encoder enhanced with learnable temporal position encoding extracts context-aware embeddings from multi-period financial sequences, and time-weighted average pooling is used to generate time-series summaries. These summaries are concatenated with static control variables and input into XGBoost for risk classification. Finally, Kernel SHAP combined with a two-stage mapping mechanism is used to trace the predictive contribution back to the original financial indicators and their corresponding historical time steps, achieving fine-grained dynamic interpretation. Experimental results show that the proposed model achieves accuracy (0.891), precision (0.826), and recall (0.723), respectively. The Area Under Curve (AUC) value for the Receiver Operating Characteristic Curve (ROC) is 0.809, and the SHAP values are well-distributed with few extreme values. The conclusions show that using the model presented in this paper helps improve the accuracy of financial risk prediction and provides reliable interpretability support for corporate risk management.