CSSFormer-VaR: A Cross-Scale State-Space Deep Learning Framework for Robust Financial Value-at-Risk Forecasting
TL;DR
The model introduces a Mamba-Enhanced Multi-Scale Bidirectional Fusion module (MEMBA-F) and a Cross-Scale Contextual Attention Gating module (CSCAG) to jointly model global temporal dependencies and local fluctuation features, thereby enhancing the identification capability for extreme risk events.
Abstract
This paper proposes a Value-at-Risk (VaR) forecasting model based on an improved Transformer architecture to address the nonlinear and multi-scale dependency characteristics in financial market risk prediction. The model introduces a Mamba-Enhanced Multi-Scale Bidirectional Fusion module (MEMBA-F) and a Cross-Scale Contextual Attention Gating module (CSCAG) to jointly model global temporal dependencies and local fluctuation features, thereby enhancing the identification capability for extreme risk events. Systematic experiments were conducted on seven U.S.-listed stock datasets, including Alibaba, Amazon, GOOGL, MSFT, TSLA, Twitter, and Visa. The results show that, compared with mainstream methods such as LSTM, BiLSTM, Transformer, and TimeMixer, the proposed model achieved the highest average $R^{2}$ value of 0.9034 and the lowest average MAE value of 0.0020 in VaR forecasting tasks, demonstrating strong fitting ability and prediction stability. Furthermore, noise perturbation and rolling prediction experiments further verified the robustness and generalization performance of the model under highly volatile market conditions. This study provides a high-precision, structure-aware, and generalizable solution for intelligent modeling of complex financial time-series risks.