ReSGA beats 12 benchmarks in VaR and Expected Shortfall forecasting
A million-parameter model outperforms 12 rivals on 97 years of US stock data...
A new paper from Yichi Zhang, Ke Zhu, and Zhoufan introduces ReSGA (retrieval-enhanced self-grouping autoencoder), a large tail risk model designed for learning Value-at-Risk (VaR) and Expected Shortfall (ES). Traditional approaches with limited parameters suffer from model misspecification in the big data era. ReSGA addresses this by leveraging millions of parameters to capture rich cross-sectional dependence and long-term temporal dynamics of assets using 153 firm characteristics. The model is trained on monthly US equity returns spanning 1926 to 2023.
ReSGA outperforms twelve econometric and machine learning competitors across out-of-sample loss and statistical backtesting. Its forecast advantages translate into significant economic gains from long-short decile portfolios constructed with a new size-enhanced left-side momentum strategy. A systematic scaling analysis reveals that improvements in joint VaR-ES forecasting are primarily driven by data complexity rather than model complexity. Additionally, group-importance and transfer-learning analyses demonstrate the model's interpretability and cross-market generalizability.
- ReSGA uses millions of parameters to capture cross-sectional and temporal dynamics from 153 firm characteristics.
- Outperformed 12 econometric and ML models on 97 years of US equity data for VaR and ES forecasting.
- Forecast gains yield economic benefits via a size-enhanced left-side momentum long-short portfolio strategy.
Why It Matters
Transforms risk management by using big data and deep learning for more accurate tail risk forecasts.