A Deep Learning Approach for Trading Factor Residuals
A Deep Learning Approach for Trading Factor Residuals
This paper replicates the Deep Learning Statistical Arbitrage (DLSA) methodology of Guijarro-Ordonez et al. (2019) on U.S. equity data from 2016 to 2024, strictly adhering to point-in-time principles. The approach uses factor models (Fama-French 5-factor, PCA, IPCA) to isolate residual returns, then applies a CNN-Transformer architecture to detect exploitable patterns. The trading strategy aims to maximize out-of-sample Sharpe ratios with constraints on leverage and turnover. Surprisingly, the replication yields out-of-sample Sharpe ratios occasionally exceeding 10, far higher than the original study's ~3.5–4. While this suggests strong predictive power, the authors caution that such extreme results may stem from model overfitting, unique market conditions (e.g., post-COVID dynamics), or the omission of transaction costs and market impact. The paper emphasizes that further robustness checks, including realistic friction modeling and testing on alternative periods, are necessary before drawing definitive conclusions about the strategy's economic viability.
Highlights
- 1Replicates the Deep Learning Statistical Arbitrage (DLSA) methodology on a more recent out-of-sample period (2016–2024) with strict point-in-time data handling.
- 2Observes out-of-sample Sharpe ratios occasionally exceeding 10, significantly higher than the original study's ~3.5–4.
- 3Employs CNN + Transformer architectures to detect patterns in factor model residuals for statistical arbitrage.
- 4Highlights potential overfitting, market regime changes, and omitted transaction costs as cautionary factors for the extreme results.
- 5Emphasizes the need for robustness checks and realistic trading frictions before concluding economic significance.
Methods
- MFactor models (Fama-French 5-factor, PCA, IPCA) to compute residual returns from systematic risk components.
- MConvolutional Neural Network (CNN) for local pattern extraction from residual time series.
- MTransformer with attention mechanism for capturing global temporal dependencies between local patterns.
- MRolling window training (1000 days) and out-of-sample testing (125 days) with feedforward network for portfolio weight optimization.
Results
- ROut-of-sample Sharpe ratios occasionally exceed 10, far surpassing the original study's ~3.5–4.
- RPreliminary backtests show high cumulative returns and favorable performance metrics, but without transaction costs.
- RResults raise concerns about model overfitting and sensitivity to the specific 2016–2024 market regime.
- RStrict point-in-time data handling and no information leakage were confirmed, yet extreme performance persists.
- RFurther robustness checks and incorporation of market frictions are required to validate the findings.
Analyze Paper
Generate insights from "A Deep Learning Approach for Trading Factor Residuals".