Title: ReSGA: A High-Complexity Tail Risk Framework for Estimating Value-at-Risk and Expected Shortfall

Abstract: Accurately estimating Value-at-Risk (VaR) and Expected Shortfall (ES) is fundamental to effective financial risk management. However, traditional methods with constrained parameter sets are increasingly susceptible to model misspecification, particularly within the context of big data. To overcome these shortcomings, we introduce ReSGA, a retrieval-enhanced self-grouping autoencoder that functions as a large tail risk model. This architecture, comprising millions of parameters, leverages asset characteristics to capture extensive cross-sectional dependencies and long-term temporal dynamics. In an empirical study utilizing monthly US equity returns spanning from 1926 to 2023 and incorporating 153 firm-specific features, ReSGA demonstrated superior performance compared to twelve competing econometric and machine learning models, achieving lower out-of-sample loss and passing statistical backtests. Furthermore, the predictive edge provided by ReSGA yields substantial economic benefits, as evidenced by long-short decile portfolios generated through a novel size-enhanced left-side momentum strategy. To isolate the impact of complexity, we performed a systematic scaling analysis revealing that enhancements in joint VaR-ES forecasting stem mainly from the complexity of the data rather than the model’s structural complexity. Lastly, investigations into group importance and transfer learning underscore the model’s interpretability and its ability to generalize across different markets.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC