Multi-Source Data-Driven Dynamic Financial Health Evaluation for Enterprises

This study developed a dynamic financial health evaluation model that integrated multi-source incomplete views and exponential time decay weights. The framework addressed critical bottlenecks in traditional assessments, including decision lag from static statements, systematic bias caused by missing heterogeneous data, and “concept drift” where old samples failed to reflect current market cycles. The model fused financial ratios, macro indicators, and Natural Language Processing-extracted sentiment. It employed variational inference with masking to reconstruct missing features, Least Absolute Shrinkage and Selection Operator for dimensionality reduction, and a gradient boosting decision tree adjusted for temporal relevance. Empirical tests on A-share manufacturing data yielded 91.2% accuracy and an Area Under Curve (AUC) of 0.935. Notably, the model maintained high robustness (0.852 AUC) despite a 40% missing data rate. These results signified a paradigm shift toward proactive risk management, offering a rigorous tool for early risk intervention.