DFRAEWM: A Deep Learning Model with Self-Attention and GRU for Digital Financial Risk Assessment and Early Warning

Abstract

Under the wave of digitalization, corporate financial data has grown massively, and traditional financial risk assessment and early warning models are difficult to cope with. This study constructs a deep financial risk assessment and early warning model (DFRAEWM), integrating self-attention mechanism, GRU and other technologies. The experiment uses real financial data, covering 20 companies and information from 2020 to 2024 , and compares models such as linear discriminant analysis (LDA) and shallow neural network (SNN). The results show that DFRAEWM far exceeds traditional and simple deep learning models in terms of accuracy (0.92), precision (0.90), recall (0.91), F1 score (0.905) and AUC-ROC (0.95). The advantages are significant in different industries (such as the manufacturing industry has an accuracy of 0.93), enterprise scale (large enterprises have an accuracy of 0.91) and time span (5-year recall rate of 0.93). Research shows that DFRAEWM can effectively explore financial characteristics, capture risk evolution, and provide enterprises with reliable risk assessment and early warning, which is of great significance to promoting the digital transformation of corporate financial management.This study uses a dataset collected between January 1, 2020 and December 31, 2024 to develop and evaluate the proposed models. All analyses, experiments, and reported results are therefore based on records within this 2020–2024 interval. Any prior references in the manuscript to data from 2025 were typographical errors and have been replaced with the consistent 2020–2024 range.The proposed DFRAEWM model was evaluated on a real-world financial dataset covering 2020–2024. The results show that DFRAEWM consistently outperforms baseline models, including SVM and Logistic Regression, in terms of accuracy, F1-score, and robustness across multiple test scenarios. In particular, DFRAEWM achieved a 12% improvement in accuracy and an 8% gain in F1 compared with the best-performing baseline. These results demonstrate the model’s ability to effectively capture temporal dependencies and feature importance for financial trend forecasting.