Paper: arXiv 2503.20987
Authors: Songci Xu, Qiangqiang Cheng, Chi-Guhn Lee
Abstract
In the realm of stock prediction, machine learning models encounter considerable obstacles due to the inherent low signal-to-noise ratio and the nonstationary nature of financial markets. These challenges often result in spurious correlations and unstable predictive relationships, leading to poor performance of models when applied to out-of-sample (OOS) domains. To address these issues, we investigate \textit{“Domain Generalization”} techniques, with a particular focus on causal representation learning to improve a prediction model’s generalizability to OOS domains. By leveraging multi-factor models from econometrics, we introduce a novel error bound that explicitly incorporates causal relationships. In addition, we present the connection between the proposed error bound and market nonstationarity. We also develop a \textit{“Causal Discovery”} technique to discover invariant feature representations, which effectively mitigates the proposed error bound, and the influence of spurious correlations on causal discovery is rigorously examined. Our theoretical findings are substantiated by numerical results, showcasing the effectiveness of our approach in enhancing the generalizability of stock prediction models.
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 involves advanced mathematics with formal error bounds, Wasserstein distances, and causal derivation (High Math). It validates theory with numerical experiments and discusses implementation details like domain definition, though it lacks full backtest results or code (Moderate-High Rigor).
Research Flowchart
flowchart TD
A["Research Goal: Improve Out-of-Sample Generalizability of Stock Prediction Models"] --> B{"Methodology"}
B --> C["Domain Generalization & Causal Representation Learning"]
B --> D["Causal Discovery for Invariant Feature Representations"]
C --> E["Theoretical Framework"]
D --> E
E --> F["Novel Error Bound<br>incorporating causal relationships & market nonstationarity"]
F --> G["Experiments: Mitigating Spurious Correlations"]
G --> H["Outcome: Enhanced Generalizability & Stable Predictions OOS"]
E --> H
classDef methodology fill:#e1f5fe,stroke:#01579b,stroke-width:2px;
class C,D,E methodology;
classDef outcome fill:#e8f5e8,stroke:#2e7d32,stroke-width:2px;
class H outcome;
subgraph Data_Inputs ["Data & Inputs"]
direction LR
I1["Financial Market Data"]
I2["Multi-Factor Models"]
end
Data_Inputs --> B
B --> Data_Inputs
classDef data fill:#fff3e0,stroke:#e65100,stroke-width:1px;
class I1,I2 data;
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 a novel error bound that explicitly incorporates causality from multi‑factor models and introduces a Causal Discovery technique to learn invariant feature representations that reduce this error bound.
Source sentence
ii. We develop the Causal Discovery technique to learn invariant feature representations that reduces the proposed error bound, whose influence from spurious correlations are also rigorously analyzed.
Assumes 1
- For all T>0 and τ>0, the minima of the set {P(JT′ > τ) | T′ ≥T} exist. other
Source sentence
Assumption 3.1: ∀T > 0 and ∀τ > 0, the minima of the set {P(JT ′ > τ) | T ′ ≥T} exist.
Data
2022-09 – 2024-03 · stock universe · equities · daily
- Chinese A‑share market market-prices public
Source sentence
We present a comprehensive analysis of the Chinese A‑share market over the period from January 4, 2010, to March 4, 2024, encompassing more than 4000 stocks.
- Alpha191 dataset alternative commercial
Source sentence
The Alpha191 dataset, provided by GUOTAI JUNAN SECURITIES, consists of 191 quantitative 1‑dimensional signals tailored for the Chinese market, each representing a distinct formula based on various market data such as prices and volumes.
Costs
Not modeled
Finds 3
Sharpe 2.04 · annual return 7.2%
- Causal discovery algorithms (˜K ≥1) consistently outperform baseline models in both Sharpe Ratio and CAGR across all architectures and sub‑periods. performance
Source sentence
Regardless of architectures or subperiods, the causal discovery algorithms (i.e., ˜K ≥1) outperform the baseline algorithms in terms of both SR and CAGR.
- Increasing the number of causal discovery components improves performance, with ˜K = 4 achieving the highest SR (2.04) and CAGR (7.22%). performance
Source sentence
˜K = 1: SR = 1.64, CAGR = 6.04%; ˜K = 2: SR = 1.91, CAGR = 6.88%; ˜K = 3: SR = 1.88, CAGR = 6.66%; ˜K = 4: SR = 2.04, CAGR = 7.22%.
- Ensembling model outputs over multiple runs and architectures shows that ˜K ≥2 significantly outperforms ˜K = 1, and ˜K = 4 yields the best Sharpe Ratio and CAGR. performance
Source sentence
Figure 1 shows the equity curves induced by the ensembled outputs, demonstrating that ˜K ≥2 significantly outperforms ˜K = 1, and ˜K = 4 results in the best SR and CAGR.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 25 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.