RiskMamba

To tackle the limitations of traditional time-series models in capturing long-term dependencies, managing high computational costs, and adapting to volatile non-stationary financial data, this study introduces RiskMamba, a lightweight enterprise financial risk prediction model. It integrates Hierarchical Residual Multi-Scale Normalization (HRMS Norm) for stable gradient flow, Temporal Convolutional Networks (TCN) with dilated convolution to expand receptive fields for long-term dependency analysis, and the Mamba-2 Block to merge state-space modeling with parallel convolution for non-stationary sequence processing. Experiments on four datasets (Give Me Some Credit, S&P 500, Cryptocurrency Market, European Banking Stress Test) show RiskMamba outperforms baselines, achieving up to 15.7% lower MSE. By addressing non-stationarity and long-term dependency challenges, it offers a scalable framework for organizational risk management, enabling efficient edge computing deployment.