Paper: arXiv 2402.08233

Authors: Fabian Krause, Jan-Peter Calliess

Abstract

In Statistical Arbitrage (StatArb), classical mean reversion trading strategies typically hinge on asset-pricing or PCA based models to identify the mean of a synthetic asset. Once such a (linear) model is identified, a separate mean reversion strategy is then devised to generate a trading signal. With a view of generalising such an approach and turning it truly data-driven, we study the utility of Autoencoder architectures in StatArb. As a first approach, we employ a standard Autoencoder trained on US stock returns to derive trading strategies based on the Ornstein-Uhlenbeck (OU) process. To further enhance this model, we take a policy-learning approach and embed the Autoencoder network into a neural network representation of a space of portfolio trading policies. This integration outputs portfolio allocations directly and is end-to-end trainable by backpropagation of the risk-adjusted returns of the neural policy. Our findings demonstrate that this innovative end-to-end policy learning approach not only simplifies the strategy development process, but also yields superior gross returns over its competitors illustrating the potential of end-to-end training over classical two-stage approaches.

Complexity vs Empirical Score

  • Math Complexity: 8.0/10
  • Empirical Rigor: 6.0/10 (capped from 6.5: transaction costs not modelled; see penalty flags)
  • Quadrant: Holy Grail — high math complexity, high empirical rigor (effective empirical rigor capped from 6.5)

Why this score: The paper employs advanced concepts like policy gradient, end-to-end training of neural networks, and references stochastic calculus and OU processes, indicating high math complexity. It includes backtesting on historical US equity returns and comparisons to benchmarks, demonstrating significant empirical rigor.

Research Flowchart

  flowchart TD
  A["Research Goal: <br> Generalise StatArb with Data-Driven Methods"] --> B["Input: US Stock Returns Data"]
  B --> C["Phase 1: Classical Autoencoder <br> (Unsupervised Feature Extraction)"]
  C --> D["Phase 2: Policy Learning <br> (End-to-End Neural Network)"]
  D --> E["Computational Process: <br> Backpropagation of Risk-Adjusted Returns"]
  E --> F["Outcome: Direct Portfolio Allocations"]
  F --> G["Key Finding: <br> Superior Gross Returns vs. Competitors"]

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 embedding an Autoencoder into a neural network that directly outputs portfolio allocations and is trained end‑to‑end on risk‑adjusted returns.

Source sentence
We propose embedding an Autoencoder in a neural network representing a trading policy, trained end-to-end with a loss function that optimises both portfolio return representation and risk‑adjusted returns.

Assumes 5

  • Asset pricing models rest on the assumption of no arbitrage. model
    Source sentence
    Asset pricing models rest on the assumption of no arbitrage - the premise that all excess returns are attributable to risk factors.
  • Weights computed at time t can be used to take positions at the beginning of time t+1. execution
    Source sentence
    We assume we can take positions using this weight at the beginning of time t + 1 and capture the return from the beginning to the end of t+1.
  • Target constant leverage of 1 for the portfolio. execution
    Source sentence
    The simplest combination, and the one we use throughout to highlight and compare the modelling aspects of a statistical arbitrage trading pipeline, is to target constant leverage in the portfolio, i.e. wi,t = φi(ǫi,t−τ:t) / ||φi(ǫi,t−τ:t)||1.
  • The considered universe of stocks is liquid enough for daily turnover. market
    Source sentence
    This assumption is not unrealistic given the liquidity of the considered universe of stocks and the state of electronic markets.
Show 1 more
  • Universe filtered to exclude low‑price, small‑cap, low‑volume stocks. data
    Source sentence
    We restrict our universe by removing any stocks with a close price lower than $5, a 20‑day median rolling market capitalisation of below $1bn, or a 20‑day

Data

  • CRSP stock data · CRSP market-prices public
    January 2000 to December 2022 · US traded companies, primary listing
    Source sentence
    stock data was used in the experiments from January 2000 to December 2022.
  • Fama French factor returns · Fama French benchmark public
    January 2000 to December 2022
    Source sentence
    For our benchmark analysis, we use Fama French factor returns for the same time period.

Costs

Not modeled A trading or execution result is reported without costs.

Source sentence
Our end‑to‑end Autoencoder architecture of learning the entire pipeline in StatArb modelling shows promising results on a pre‑cost basis.

Finds 4

  • Basic Autoencoder models achieve competitive returns but do not surpass PCA in Sharpe ratio. performance
    Source sentence
    while the Autoencoder models do not surpass PCA in terms of Sharpe ratio, the regression‑based models deliver competitive, and in some cases superior, returns when compared to the best‑performing PCA models
  • The optimal number of latent factors for the policy aligns with literature recommendations of 10‑15. theoretical
    Source sentence
    the number of latent factors in the best‑performing policy models aligns with the literature’s recommended range of 10 to 15
  • The end‑to‑end policy learning approach simplifies strategy development and yields superior gross returns over competitors. performance
    Source sentence
    Our findings demonstrate that this innovative end‑to‑end policy learning approach not only simplifies the strategy development process, but also yields superior gross returns over its competitors
  • Basic Autoencoder implementations can rival established benchmark models like PCA and Fama French in a StatArb context. performance
    Source sentence
    even basic Autoencoder implementations can rival the performance of established benchmark models like PCA and Fama French in a StatArb context

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 21 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 2 papers in the archive (2 by independent authors)

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.