Attention Factors for Statistical Arbitrage

Battle Count: 71

This paper is directly and fundamentally about quantitative trading. It addresses the core problem of statistical arbitrage: identifying similar assets, detecting mispricing, and constructing profitable trading strategies after transaction costs. The reported net Sharpe ratio of 2.3 on US equities over 24 years is exceptional and represents a significant advance over prior work (84% improvement over state-of-the-art). The one-step joint optimization of factor construction and trading policy with explicit transaction cost modeling is highly relevant for practitioners. The interpretable factor structure (industry sectors) and the finding that weak factors matter for arbitrage are practically valuable insights.

The model involves several non-trivial components: (1) attention mechanism over cross-sectional firm characteristics with learned embeddings, (2) LongConv sequence model with FFT-based convolution, kernel regularization (Squash operator), and geometric decay initialization, (3) joint optimization of factor weights and trading policy with a composite objective (net Sharpe + explained variance), (4) transaction cost modeling within the loss function, (5) rolling window retraining. The architecture is moderate in size (hidden dim 32, 1 layer LongConv, 30 epochs), but the joint optimization with the Sharpe ratio objective and the attention mechanism over the cross-section require careful implementation. The paper provides detailed hyperparameters but no code.

The paper provides detailed hyperparameters (Table 4), model architecture descriptions, data sources (CRSP/Compustat), and implementation details (Adam optimizer, 5 NVIDIA RTX A6000 GPUs, CUDA 12.5). However, no GitHub repository or code link is provided. The firm characteristics construction references the Internet Appendix of Chen, Pelger, and Zhu (2024). The transaction cost model is specified (5 bps per transaction, 1 bps shorting cost). Rolling window estimation (8-year training, annual retraining) is clearly described. Reproduction would require access to CRSP/Compustat databases and careful implementation of the attention factor and LongConv components.