Paper: arXiv 2510.11616

Authors: Elliot L. Epstein, Rose Wang, Jaewon Choi, Markus Pelger

Abstract

Statistical arbitrage exploits temporal price differences between similar assets. We develop a framework to jointly identify similar assets through factors, identify mispricing and form a trading policy that maximizes risk-adjusted performance after trading costs. Our Attention Factors are conditional latent factors that are the most useful for arbitrage trading. They are learned from firm characteristic embeddings that allow for complex interactions. We identify time-series signals from the residual portfolios of our factors with a general sequence model. Estimating factors and the arbitrage trading strategy jointly is crucial to maximize profitability after trading costs. In a comprehensive empirical study we show that our Attention Factor model achieves an out-of-sample Sharpe ratio above 4 on the largest U.S. equities over a 24-year period. Our one-step solution yields an unprecedented Sharpe ratio of 2.3 net of transaction costs. We show that weak factors are important for arbitrage trading.

Complexity vs Empirical Score

  • Math Complexity: 7.0/10
  • Empirical Rigor: 6.0/10 (capped from 8.5: reported Sharpe ratio of 3.97 is implausibly high for daily or lower-frequency data; see penalty flags)
  • Quadrant: Holy Grail — high math complexity, high empirical rigor (effective empirical rigor capped from 8.5)

Why this score: The paper employs advanced mathematical techniques including deep learning with attention mechanisms, latent factor models, and sequence models (e.g., transformers, S4), indicating high mathematical complexity. It presents a comprehensive empirical study with a 24-year backtest on U.S. equities, reporting specific Sharpe ratios (gross and net), transaction costs, and out-of-sample results, demonstrating strong empirical rigor.

Research Flowchart

  flowchart TD
  A["Research Goal: Develop a Joint Framework for Statistical Arbitrage"] --> B["Methodology: Attention Factor Model"]
  B --> C["Data: Firm Characteristic Embeddings"]
  C --> D{"Computational Process"}
  D --> E["Learn Conditional Latent Attention Factors"]
  E --> F["Model Residual Portfolios with General Sequence Model"]
  F --> G["Joint Estimation of Factors & Trading Policy"]
  G --> H["Key Outcomes"]
  H --> I["Out-of-sample Sharpe Ratio > 4.0"]
  H --> J["Net Sharpe Ratio 2.3 (Transaction Costs)"]
  H --> K["Weak Factors are Crucial for Arbitrage"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes an Attention Factor Model that jointly learns tradable arbitrage factors and portfolio allocations in a one‑step framework optimized for trading costs.

Source sentence
In this paper, we propose the Attention Factor Model, a framework that jointly learns tradable arbitrage factors and arbitrage portfolio allocations in a computationally efficient manner.

Assumes 2

  • Asset returns follow a conditional factor model with time‑varying loadings based on lagged information. model
    Source sentence
    We assume that asset returns can be modeled by a conditional factor model: 𝑅𝑖,𝑡= 𝛽𝑇𝑖,𝑡−1𝐹𝑡+ 𝜖𝑖,𝑡,
  • Only the 500 most liquid U.S. equities (largest by market cap) are used in the empirical analysis. data
    Source sentence
    We consider only the 500 largest stocks based on the previous month market capitalization.

Data

Backtest sample: 1998 – 2021 · U.S. equities · equities · daily

  • CRSP daily equity returns · CRSP market-prices public
    January 1990 through December 2021 · 500 largest U.S. stocks
    Source sentence
    We collect daily equity return data for the securities on CRSP from January 1990 through December 2021.
  • Firm-specific characteristics · CRSP/Compustat fundamentals public
    Source sentence
    ... All these variables are constructed either from accounting variables from the CRSP/Compustat database or from past returns from CRSP.

Recognised sources: CRSP Compustat

Costs

Modeled Costs are part of the reported headline results.

Cost = 0.0005 per unit turnover (5 bps) plus 0.0001 per unit short position (1 bp). Net return = gross return minus this cost.

Source sentence
... The portfolio net return is then calculated as: 𝑅port𝑡,𝑛𝑒𝑡= 𝑅port𝑡−cost(𝜔𝑡,𝜔𝑡−1).

Finds 4

Reported performance (Attention Factors): Sharpe 3.97 · annual return 16.7%

  • The Attention Factor model achieves an out‑of‑sample Sharpe ratio above 4 before costs and a net Sharpe of 2.3 after transaction costs. performance
    Source sentence
    Our Attention Factor model provides unprecedented performance of 2.3 after trading costs.
  • Simple PCA factor models suffer from high turnover, yielding a net Sharpe around 1.5 after costs. performance
    Source sentence
    a simple PCA factor model cannot adjust the factor construction to trading frictions, resulting in a deterioration of the net performance to a Sharpe ratio of around 1.5.
  • Including weak factors improves performance; using 30 attention factors outperforms using only 8. performance
    Source sentence
    We observe an out‑of‑sample improvement for including 30 factors. These higher order factors capture weak signals and local dependency patterns.
  • The model learns economically meaningful factor loadings that cluster firms by industry without using explicit industry labels. other
    Source sentence
    Firms in similar industries are grouped together, that is, our model learns specific industries. For example, the upper right cluster represents banks and financial ...

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 30 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (64)

  1. Attention is All you Need (2017) — link
  2. Conditional Latent Factor Models Via Econometrics-Based Neural Networks (2021) — link
  3. Large covariance estimation by thresholding principal orthogonal complements (2013) — link
  4. Determining the Number of Factors in Approximate Factor Models (2002) — link
  5. Predicting Corporate Bond Returns: Merton Meets Machine Learning (2022) — link
  6. Dissecting characteristics nonparametrically (2020) — link
  7. Bond risk premia with machine learning (2021) — link
  8. Empirical Asset Pricing Via Machine Learning (2020) — link
and 56 more
  1. Forest through the Trees: Building Cross-Sections of Stock Returns (2023) — link
  2. Optimal mean reversion trading with transaction costs and stop-loss exit (2015) — link
  3. Model-based pairs trading in the bitcoin markets (2016) — link
  4. Algorithmic trading of co-integrated assets (2016) — link
  5. Optimal pairs trading: A stochastic control approach (2008) — link
  6. Dynamic portfolio selection in arbitrage (2007) — link
  7. Identifying small mean-reverting portfolios (2011) — link
  8. Risk control of mean-reversion time in statistical arbitrage (2017) — link
  9. Statistical arbitrage in the US equities market (2010) — link
  10. The profitability of pairs trading strategies: distance, cointegration and copula methods (2016) — link
  11. Pairs Trading: Quantitative Methods and Analysis (2004)
  12. Machine-learning the skill of mutual fund managers (2023) — link
  13. Understanding Systematic Risk: A High-Frequency Approach (2020) — link
  14. State-Varying Factor Models of Large Dimensions (2022) — link
  15. Factors That Fit the Time Series and Cross-Section of Stock Returns (2020) — link
  16. Interpretable Sparse Proximate Factors for Large Dimensions (2022) — link
  17. Deep Learning in Asset Pricing (2024) — link
  18. Simple Hardware-Efficient Long Convolutions for Sequence Modeling (2023) — link
  19. A Set-Sequence Model for Time Series — → 2505.11243
  20. Deep Learning Statistical Arbitrage (2025) — link
  21. Characteristics are covariances: A unified model of risk and return (2019) — link
  22. Principal Component Analysis (2002) — link
  23. Common risk factors in the returns on stocks and bonds (1993) — link
  24. Conditional Latent Factor Models Via Econometrics-Based Neural Networks (2021)
  25. State-Varying Factor Models of Large Dimensions (2020) — link
  26. APT or ``AIPT’’? The Surprising Dominance of Large Factor Models (2024) — link
  27. Estimating latent asset-pricing factors (2020) — link
  28. Artificial Intelligence Asset Pricing Models (2025) — link
  29. Understanding Systematic Risk: A High-Frequency Approach (2020) — link
  30. Statistical Arbitrage in the U . S . Equities Market (2008)
  31. Flexible least squares for temporal data mining and statistical arbitrage (2009) — link
  32. Risk control of mean-reversion time in statistical arbitrage (2017) — link
  33. Mispricing Factors (2016) — link
  34. Pairs Trading (2005) — link
  35. Developing Arbitrage Strategy in High-frequency Pairs Trading with Filterbank CNN Algorithm (2018) — link
  36. Identifying small mean-reverting portfolios (2011)
  37. Statistical Arbitrage in Cryptocurrency Markets (2019) — link
  38. Pairs Trading: Performance of a Relative-Value Arbitrage Rule (2006) — link
  39. Crossformer: Transformer Utilizing Cross-Dimension Dependency for Multivariate Time Series Forecasting (2023) — link
  40. iTransformer : Inverted Transformers Are Effective for Time Series Forecasting (2024) — link
  41. Efficiently Modeling Long Sequences with Structured State Spaces (2022) — link
  42. Portfolio Transformer for Attention - Based Asset Allocation (2022) — link
  43. Autoencoder asset pricing models (2021) — link
  44. Stockformer: A Price-Volume Factor Stock Selection Model Based on Wavelet Transform and Multi-Task Self-Attention Networks — → 2401.06139
  45. A universal end-to-end approach to portfolio optimization via deep learning (2021) — link
  46. End-to-end, decision-based, cardinality-constrained portfolio optimization (2025) — link
  47. Shrinking the cross-section (2020) — link
  48. NeuralFactors: A Novel Factor Learning Approach to Generative Modeling of Equities — → 2408.01499
  49. Forecasting using principal components from a large number of predictors (2002)
  50. Deep factor model: Explaining deep learning decisions for forecasting stock returns with layer-wise relevance propagation (2018)
  51. High-dimensional multivariate forecasting with low-rank Gaussian Copula Processes (2019) — link
  52. Think Globally, Act Locally: A Deep Neural Network Approach to High-Dimensional Time Series Forecasting (2019) — link
  53. A five-factor asset pricing model (2015) — link
  54. Decoupled weight decay regularization (2019) — link
  55. Visualizing Data using t-SNE (2008) — link
  56. Deep risk model: a deep learning solution for mining latent risk factors to improve covariance matrix estimation (2022) — link

3 of 64 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.