On statistical arbitrage under a conditional factor model of equity returns

Trent Spears, Stefan Zohren, Stephen RobertsarXiv 2023 · DOI · arXiv

TL;DR

This paper introduces a state-space framework for statistical arbitrage that models equity returns using a conditional factor model. By employing Kalman Filters to estimate latent, time-varying risk premia, the authors identify "fair value" dislocations for mean-reversion trading, achieving competitive performance and robust Sharpe ratios over a comprehensive 29-year backtest of US equities.

What Problem Does This Paper Address?

Statistical arbitrage (StatArb) is built on the premise that asset prices eventually revert to a "fair value." The central challenge lies in accurately defining and estimating that fair value in a market that is inherently non-stationary—meaning its statistical properties change over time.

While many existing strategies rely on simple pairs trading or Principal Component Analysis (PCA), these often struggle to scale to large portfolios or fail to account for the dynamic nature of market risk factors. This paper seeks to bridge that gap by providing a mathematically rigorous way to estimate "fair" returns across a multivariate portfolio, specifically accounting for how the rewards for different risks (like momentum or value) shift over time.

Methods

The authors utilize a state-space framework to model the relationship between asset returns and risk factors.

  • The Model: They assume asset returns are "noisy observations" of a linear combination of known factor exposures (e.g., how sensitive a stock is to market volatility) and latent (hidden) factor risk premia.
  • Linear Approach: They use the classic Kalman Filter (KF) to estimate the hidden state of these risk premia in an "online" fashion (updating as new data arrives).
  • Non-linear Approach: They implement an Unscented Kalman Filter (UKF), where the state transition is governed by a Multi-Layer Perceptron (MLP)—a type of neural network. This allows the model to potentially capture complex, non-linear evolutions in market risk.
  • Trading Strategy: The model generates "filtered" returns. When the actual observed return deviates significantly from this filtered fair value (measured by a Z-score), the system triggers a mean-reversion trade (longing the undervalued or shorting the overvalued asset).

Datasets

The study uses a massive dataset of US equity market data spanning 29 years (1992–2021).

  • Source: CRSP and IBES databases via the Wharton Research Data Service (WRDS).
  • Universe: The top 2,000 US stocks by market capitalization on any given trading day.
  • Factors: Five specific risk factors were constructed for each asset: Market Beta, Size, Volatility, Momentum, and Value.
  • Benchmark: The strategy was compared against the S&P 500 (excess returns) and a simple rolling-mean reversal benchmark.

Key Insights and Ideas

  • Capturing Non-Stationarity: By modeling risk premia as time-varying latent variables rather than static constants, the model "de facto" captures the changing regimes of the stock market.
  • The Complexity Paradox: Interestingly, the non-linear UKF model did not significantly outperform the linear KF model. This suggests that for daily frequency StatArb, the added computational complexity of neural networks for state transition might not yet be justified by the available signal-to-noise ratio.
  • Diversification through Blending: The authors found that "blending" the StatArb strategy with a standard long-only market portfolio significantly improved the Sharpe ratio. Because the StatArb strategy is beta-neutral (hedged against market swings), it provides a diversifying return stream that complements traditional buy-and-hold investing.

Key Results

  • Strong Historical Performance: The primary model achieved an average annual Sharpe ratio of 1.29 over the full 1993–2021 period.
  • Competitive Edge: The results were competitive with, and in some cases superior to, other cutting-edge methods in the literature, such as those based on PCA or Vine Copulas.
  • Transaction Cost Robustness: The strategy remained profitable after accounting for realistic transaction costs (5 basis points per trade), though performance decayed as costs approached 15 bps.
  • Modern Challenges: Performance was notably stronger in the first half of the dataset (1993–2007). In more recent years (2015–2019), the "alpha" or excess return diminished, likely due to increased market efficiency and high-frequency competition.

Limitations and Future Work

  • Alpha Decay: The authors acknowledge a noticeable degradation in strategy performance in the most recent years of the study, suggesting that the "low-hanging fruit" of daily mean reversion is being picked by more participants.
  • Data Granularity: The study focused on daily data. The authors suggest that moving to intraday timeframes might reveal more significant non-linearities where the UKF could shine.
  • Alternative Data: The current model relies purely on price-derived factors. Incorporating alternative data (sentiment, supply chain data, etc.) could improve the predictive power of the state-space model.
  • Regime Switching: Future work could involve adding "regime filters" to help the model automatically adjust its trading aggressiveness based on whether the market is in a high-volatility or low-volatility state.

Summary updated

Abstract

We consider a conditional factor model for a multivariate portfolio of United States equities in the context of analysing a statistical arbitrage trading strategy. A state space framework underlies the factor model whereby asset returns are assumed to be a noisy observation of a linear combination of factor values and latent factor risk premia. Filter and state prediction estimates for the risk premia are retrieved in an online way. Such estimates induce filtered asset returns that can be compared to measurement observations, with large deviations representing candidate mean reversion trades. Further, in that the risk premia are modelled as time-varying quantities, non-stationarity in returns is de facto captured. We study an empirical trading strategy respectful of transaction costs, and demonstrate performance over a long history of 29 years, for both a linear and a non-linear state space model. Our results show that the model is competitive relative to the results of other methods, including simple benchmarks and other cutting-edge approaches as published in the literature. Also of note, while strategy performance degradation is noticed through time -- especially for the most recent years -- the strategy continues to offer compelling economics, and has scope for further advancement.
BibTeX
@misc{spears2023statistical,
  author = {Trent Spears and Stefan Zohren and Stephen Roberts},
  title = {{On statistical arbitrage under a conditional factor model of equity returns}},
  year = {2023},
  eprint = {2309.02205},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2309.02205},
}
Figure