Study of Stock Return Prediction Based on Big Data

Shiqi Chen2023 · DOI

TL;DR

This paper explores how to distill "Big Data"—specifically 115 different fundamental indicators—into a single, powerful "quality index" for predicting stock returns in the Chinese A-share market. The author demonstrates that Partial Least Squares (PLS) significantly outperforms traditional methods like Principal Component Analysis (PCA) by specifically targeting the components of company data that correlate with future returns rather than just internal data variance.

What Problem Does This Paper Address?

In modern finance, researchers face a "factor zoo": an overwhelming number of indicators (profitability, growth, debt levels, etc.) that might influence stock prices. The challenge isn't finding data, but rather dimensionality reduction—compressing these 115 variables into a manageable signal without losing predictive power.

Furthermore, while much of this research has been conducted on US markets, this paper focuses on the Chinese stock market, testing whether fundamental analysis remains effective in a market known for different regulatory constraints (like short-selling limits) and investor behaviors.

Methods

The author categorizes 115 fundamental indicators into five dimensions: profitability, growth, corporate governance, potential value, and safety. To combine these into a "Comprehensive Quality Index," the paper compares several statistical approaches:

  • Principal Component Analysis (PCA): A standard unsupervised method that finds the directions of maximum variance in the data.
  • Fama-Macbeth Regression: A classic two-step procedure used in asset pricing to estimate risk premiums.
  • Partial Least Squares (PLS): The "star" of the paper. Unlike PCA, PLS is a supervised method; it reduces dimensionality by looking for the components that have the highest covariance with the target variable (future stock returns).
  • Univariate Portfolio Analysis: To test these indices, the author sorts stocks into ten groups (deciles) based on their quality scores and calculates the returns of a "long-short" portfolio (buying the highest quality and selling the lowest).

Datasets

The study utilizes the CSMAR database, a standard for Chinese financial research.

  • Scope: All A-shares on the Shanghai and Shenzhen Stock Exchanges.
  • Timeframe: January 1999 to December 2021.
  • Features: 115 fundamental indicators, monthly and daily returns, and standard risk factors (Fama-French factors).

Key Insights and Ideas

  • Supervised vs. Unsupervised Learning: The primary "aha moment" is the failure of PCA compared to PLS. PCA identifies what makes the data different, but those differences might be "noise" that doesn't affect stock prices. PLS specifically filters for the "fraction of returns that can be predicted," making it a far more surgical tool for investors.
  • The "Long Side" Dominance: In the Chinese market, short-selling is often restricted. The paper finds that the PLS strategy's success is driven largely by the "long side" (the high-quality stocks). This suggests the strategy is practically "viable" for retail and institutional investors who cannot easily short stocks.
  • The Small-Cap Edge: The predictive power of the quality index is stronger in smaller-cap stocks. When the portfolios were value-weighted (giving more weight to big companies), the returns decreased, suggesting that fundamental "mispricing" is more common in smaller, less-scrutinized companies.

Key Results

  • Superior Performance: The PLS-based index ($Q^{PLS}$) showed the strongest and most stable predictive power.
  • Monthly Returns: An equal-weighted long-short portfolio based on $Q^{PLS}$ generated an average monthly return of 2.07% (with a highly significant t-statistic of 4.57).
  • Risk-Adjusted Alpha: Even after accounting for market volatility, size, and value factors (using the Fama-French 3-factor model), the $Q^{PLS}$ strategy maintained a significant monthly excess return of 1.53%.
  • Market Weighting: When using market-cap weighting (more realistic for large funds), the strategy still earned a respectable 1.30% monthly return.

Limitations and Future Work

The author acknowledges that as we enter the "big data era," even linear methods like PLS might struggle with the sheer complexity of information.

  • Linearity: The models used are primarily linear. The author suggests that future research should incorporate more complex machine learning algorithms—such as LASSO or Bayesian analysis—to capture non-linear relationships between fundamentals and returns.
  • High-Dimensional Trap: While PLS helps, the "high-dimensional trap" remains a risk where models overfit to past data.
  • Market Specificity: The results are specific to the Chinese A-share market; the dynamics of the "quality factor" might shift in other emerging or developed markets.

Summary updated

Abstract

The stock market is one of the most important components of the capital market, and thus identifying the reasons for changes in expected stock returns is key to building a well-functioning capital market. Based on Gordon's dividend growth model, this paper composes company fundamental indicators fro
BibTeX
@misc{chen2023study,
  author = {Shiqi Chen},
  title = {{Study of Stock Return Prediction Based on Big Data}},
  year = {2023},
  doi = {10.4108/eai.9-12-2022.2327593},
}
Figure