Pairs Trading with Machine Learning of Similarity Factors

Can machine learning exploit many stock similarity factors to produce exceptional statistical arbitrage (pairs trading) performance? In their August 2025 paper entitled “Attention Factors for Statistical Arbitrage”, Elliot Epstein, Rose Wang, Jaewon Choi and Markus Pelger present the Attention Factor Model, which employs machine learning to:

1. Identify similar stocks based on both past returns and firm fundamentals (similarity factors).
2. Generate signals for temporary price divergences between similar stocks.
3. Set weighting/trading rules to exploit such price divergences.

Their model considers many similarity factors and the time series behaviors of these factors to maximize portfolio Sharpe ratio after transaction costs. They retrain the model each year on a rolling window of eight years of data, using the last two years of the first set of training data to select tuning parameters. We consider model variations that identify 1, 3, 5, 8, 10, 15, 30 or 100 similarity factors. They assume total costs of 0.05% one-way trading frictions and 0.01% shorting costs. They consider results of prior research as a benchmark. Using daily returns and 39 firm characteristics for the 500 largest U.S. stocks by month during January 1990 through December 2021, with model testing during January 1998 through December 2022, they find that: (more…)