Enterprise Risk Prediction Under Dynamic Distribution Shift Using Robust Adaptive Learning

Abstract

This study proposes an adversarial risk-aware robust adaptive prediction model to address the challenges of data distribution shift, adversarial perturbations, and temporal variability in enterprise risk prediction. The model consists of a feature embedding layer, an adversarially aware robust encoder, a temporal attention aggregation module, and an adaptive risk decoder. The feature embedding layer maps multi-source heterogeneous data into a unified high-dimensional latent space to represent enterprise operational, financial, and environmental features consistently. Under the constraint of the adversarial perception mechanism, the robust encoder learns stable and noise-resistant risk representations, improving reliability under noisy and anomalous conditions. The temporal attention module models dynamic temporal dependencies of risk, using time decay weights to emphasize key moments and capture the evolution of risk events. To handle cross-industry and cross-period distribution shifts, the model incorporates a distribution alignment mechanism based on maximum mean discrepancy to achieve adaptive alignment between source and target domains, enhancing generalization under different economic conditions. Experimental results show that the proposed method significantly outperforms mainstream models in accuracy, recall, and stability, achieving robust and sensitive enterprise risk identification and prediction in dynamic environments while providing a new technical pathway and theoretical foundation for intelligent risk management.