Statistical arbitrage trades temporary dislocations between similar assets. Deep learning work in this space moves from two-step pipelines — factor extraction, then signal trading — toward joint optimization that accounts for turnover and transaction costs from the start.
Two-step residual trading
The benchmark approach builds arbitrage portfolios as residuals from conditional latent factors, extracts time-series signals with a convolutional transformer, then forms a constrained trading policy.
Deep Learning Statistical Arbitrage
Residual portfolios from conditional latent factors, convolutional transformer signals, and constrained risk-adjusted trading on daily US equities. Strong gross Sharpe ratios; net performance sensitive to turnover. arXiv preprint
Joint factor and policy learning
Attention Factors learn conditional latent factors and a cost-aware trading policy in one step, so factors can adapt to reduce turnover rather than maximizing explained variance alone.
Attention Factors for Statistical Arbitrage
One-step framework with attention-based conditional factors, residual sequence signals, and a trading policy optimized for net performance. Reports Sharpe above 4 gross and 2.3 net of costs on large-cap US equities over 24 years.
Rank space representations
Indexing stocks by market-cap rank rather than name can improve mean reversion in residual portfolios. Li and Papanicolaou compare name-space and rank-space stat arb with and without neural networks.
Statistical Arbitrage in Rank Space
Builds arbitrage portfolios in rank space with stronger mean-reverting residuals than name-space equivalents. Includes an intraday rebalancing mechanism to convert between name and rank portfolios; neural nets amplify the rank-space edge.
Practical checklist
- Report gross and net Sharpe — high gross stat arb often dies after costs
- Compare two-step vs one-step designs when factor turnover is high
- Separate residual alpha from factor exposure before claiming pure arbitrage
- Walk-forward validation is mandatory; stat arb signals decay quickly
For ML validation methodology see our ML signals reading list and backtesting pitfalls guide.