ReSGA: A Large Tail Risk Model for Learning Value-at-Risk and Expected Shortfall

Battle Count: 124

Highly relevant to quantitative trading. The paper directly addresses tail risk forecasting (VaR/ES) which is critical for risk management in trading strategies. The proposed size-enhanced left-side momentum strategy demonstrates practical trading applications with significant Sharpe ratios (0.787 for ReSGA H-L portfolio) and statistically significant alphas. The cross-sectional ranking of stocks by predicted tail risk provides actionable trading signals. Transfer learning results show applicability across multiple markets. The paper bridges academic risk modeling with practical portfolio construction.

High implementation complexity due to: (1) Multi-module architecture (Encoder with GRU + group discovery, Retriever with two-stage attention, Decoder with residual connections); (2) Millions of parameters requiring GPU training; (3) Complex data preprocessing (rank-normalization, imputation) for 40,000+ stocks; (4) Expanding-window evaluation protocol; (5) FZ loss optimization with Adam; (6) Ensemble learning across multiple initializations; (7) Hyperparameter tuning across multiple scales. However, the companion website provides source code and documentation.

The paper provides a user-oriented companion website (https://tailrisk-resga.github.io) with stock-level tail risk forecasts, documentation, and links to data and source code. The dataset (Jensen, Kelly and Pedersen 2023) is publicly available. Detailed hyperparameter settings, architecture specifications, and training procedures are provided in appendices. Expanding-window evaluation protocol is clearly described.