Multiscale fuzzy wavelet-attention framework enhanced by support vector regression for intelligent financial risk and bankruptcy forecasting

Accurate forecasting of financial distress, particularly corporate bankruptcy, remains a critical challenge due to the inherent non-linearity, volatility, and uncertainty in financial time series data. Existing approaches often struggle with overfitting, a lack of interpretability, sensitivity to noise, and limited capacity to model multi-scale temporal dependencies. Furthermore, conventional machine learning models fail to capture short-term and long-term trend fluctuations while providing limited mechanisms for uncertainty quantification. This paper proposes a novel framework for intelligent financial risk forecasting: a multiscale fuzzy wavelet-attention framework with residual learning via support vector regression (MFWSVR). The proposed model integrates discrete wavelet transforms for multiscale decomposition, multi-level attention mechanisms to prioritize critical temporal patterns, and adaptive fuzzy information granules to capture uncertainty and regime shifts in financial markets. A residual learning scheme is introduced using SVR to enhance predictive precision, allowing the model to refine predictions and reduce bias in high-variance financial environments iteratively. The model under consideration has been tested on the Polish, Taiwanese, and Chinese bankruptcy prediction datasets alongside several SOTA models, such as TGNN, DeepSurv, XGBoost, and RFR. From our findings, it is evident that MFWSVR exceeds the given benchmarks in all evaluated metrics, achieving significant gains over competing models. In particular, MFWSVR obtained an MAE of 0.09 (Polish dataset), 0.11 (Taiwanese dataset), and 0.13 (Chinese dataset) at 1-year prediction horizon with R 2 values of 0.91, 0.87, and 0.83 respectively. At a 3-year horizon, the model still performs strongly with an MAE of 0.13, 0.15, and 0.17 across the three datasets, alongside high R 2 values of 0.83, 0.81, and 0.78. In contrast, baseline models TGNN and DeepSurv experience performance loss at the 3-year mark, with MAE values of 0.16 and 0.17. MFWSVR owes its ability to quantify prediction uncertainty to the integration of interval prediction through wavelet decomposition and fuzzy granulation, which is particularly valuable in financial risk management where prediction accuracy greatly influences strategic decisions.