Abstract In the modern financial system, the complexity and the connectedness of the financial market has made it more difficult to pick up the early warning signals of financial stress in companies. Most of the current statistical, machine learning and deep learning methods fail to model nonlinear relationships and long-term dependencies as well as heterogeneous financial and sentiment data at the same time. To solve these problems, a novel approach named Financial Risk Prediction framework (TCNHA-FRP) is proposed in this study, which combines a Temporal Convolutional Network with a Hierarchical Attention mechanism. The proposed workflow involves data preprocessing, DenseNet based feature extraction, modeling the temporal dependency using Temporal Convolutional Network (TCN), adaptively weighting the features using Hierarchical Attention Network (HAN) and finally making predictions in an adaptive risk-scoring layer. The TCN is trained to capture long-range temporal dependencies between firm-level and macroeconomic variables through dilated causal convolutions, and the attention mechanism is used to focus the model’s attention on the most informative number and text during prediction. The model is trained and evaluated on a multisource dataset combining financial ratios, market indicators, and sentiment-based textual disclosures. Experimental results show that TCNHA-FRP achieves 98.8% accuracy and a 98.3% F1-score, outperforming conventional machine learning methods and baseline deep learning models. The attention outputs provide interpretable insights into key risk-driving factors, supporting practical decision-making for firms, investors, and regulators. Overall, the proposed framework offers an effective and data-driven approach for timely financial risk assessment.
Wang et al. (Mon,) studied this question.