Predicting abnormal returns from news using text classification

Ronny Luss, Alexandre d’Aspremont

TL;DR

This paper demonstrates that text from news articles can predict the magnitude (but not direction) of intraday stock price movements, outperforming historical returns alone. They achieve this by using Support Vector Machines (SVMs) and a novel Multiple Kernel Learning (MKL) approach to combine text features with equity return data, leading to significantly improved prediction accuracy.

What Problem Does This Paper Address?

Predicting stock market movements has long been a central challenge in finance. Traditional models often rely on historical price data (e.g., ARCH/GARCH models) to forecast volatility, but these methods largely ignore a crucial source of information: financial news. While some studies have explored the impact of news on markets, this paper aims to systematically leverage the rich information contained in news text to predict intraday abnormal returns with greater accuracy than previously achieved. The challenge lies in effectively extracting and integrating this textual information with financial time-series data.

Methods

The core of the paper's methodology revolves around Support Vector Machines (SVMs) for classification. They frame the problem as predicting whether an "abnormal return" (defined as a price movement exceeding a certain percentile threshold) will occur within a specified time window after a news release.

Key techniques include:

  • Text Representation: News articles are converted into numerical features using a bag-of-words model, where the frequency of specific words (stems) is counted. These counts are then transformed using TF-IDF (term frequency-inverse document frequency) weighting to highlight important and discriminative words.
  • Feature Combination (Multiple Kernel Learning - MKL): To combine the predictive power of text features with historical equity return data, the authors employ Multiple Kernel Learning (MKL). MKL allows for the optimal linear combination of different "kernels" (which represent different data sources or feature types), learning the best way to weigh each source for improved prediction.
  • Analytic Center Cutting Plane Method (ACCPM): For solving the MKL problem efficiently, especially with a large number of kernels, they develop and utilize an Analytic Center Cutting Plane Method (ACCPM). This algorithm is designed to be more efficient than existing methods for large-scale MKL problems, requiring fewer SVM computations.

Datasets

  • News Data: Press releases from PRNewswire for the eight-year period from 2000-2007, focusing on publicly traded companies that issued at least 500 press releases.
  • Equity Return Data: Intraday price data from the NYSE Trade and Quote Database (TAQ) through Wharton Research Data Services.
  • Options Data: Daily options data for trading strategy evaluation, obtained from OptionMetrics.

The paper does not introduce a new benchmark dataset but rather uses existing financial data sources.

Key Insights and Ideas

  • News Predicts Magnitude, Not Direction: The study strongly suggests that news can predict the size of intraday stock returns (i.e., whether a large move will occur), but not the direction of that move. This is a crucial distinction for trading strategies.
  • Text Outperforms Historical Returns: Textual features derived from news articles consistently provide better predictive performance for abnormal returns than historical equity return data alone.
  • MKL is Key for Data Fusion: Combining different data sources (text and returns) through Multiple Kernel Learning significantly boosts prediction accuracy compared to using either source in isolation. This highlights the power of data fusion in financial forecasting.
  • ACCPM for Efficient MKL: The proposed ACCPM algorithm offers a computationally efficient way to solve large-scale MKL problems, making it practical for real-world applications.
  • Time-of-Day Effects Exist but are Less Crucial: While there's evidence of time-of-day effects on stock returns, these are less impactful for prediction than the information contained in news text or historical returns when combined via MKL.

Key Results

  • Abnormal Return Prediction: Using text features alone, the Sharpe ratio for predicting abnormal returns can reach values around 2.02 (for the LONG ONLY strategy), significantly outperforming the -1.01 Sharpe ratio achieved with absolute returns alone.
  • MKL Improvement: Combining text and absolute returns using MKL further improves performance, with accuracy reaching around 0.75 and Sharpe ratios around 16 for certain time horizons.
  • Time Horizon: Predictive power is strongest for shorter time horizons immediately following a news release and gradually decreases as the time horizon extends.
  • Threshold Sensitivity: The definition of "abnormal return" (based on percentile thresholds) significantly impacts performance, with higher thresholds (e.g., 75th or 85th percentile) generally yielding better results, especially for text-based predictions.
  • Trading Strategy Performance: A delta-hedged covered call option strategy based on MKL predictions shows a potentially large upside, particularly for predicting abnormal returns. The LONG ONLY strategy (betting on abnormal returns) performs best.

Limitations and Future Work

  • Directional Predictability: The paper reiterates that predicting the direction of returns remains elusive, consistent with existing literature.
  • Fictitious Trading Strategy: For intraday abnormal return predictions, a fictitious trading strategy was used due to the unavailability of intraday options data. Real-world trading would involve transaction costs and bid-ask spreads.
  • Feature Selection: The paper used a handpicked dictionary of words. Future work could explore more advanced feature selection techniques (e.g., RFE-SVM, latent semantic analysis) or industry-specific dictionaries.
  • Regression Analysis: Extending the work to regression analysis (Support Vector Regression - SVR) to forecast volatility and abnormal returns using MKL is suggested.
  • Tradeable Strategy: Implementing a fully tradeable strategy for daily predictions, similar to Section 2.6, is a natural next step, though intraday options data remains a hurdle.

Summary updated

Abstract

We show how text from news articles can be used to predict intraday price movements of financial assets using support vector machines. Multiple kernel learning is used to combine equity returns with text as predictive features to increase classification performance and we develop an analytic center cutting plane method to solve the kernel learning problem efficiently. We observe that while the direction of returns is not predictable using either text or returns, their size is, with text features producing significantly better performance than historical returns alone. 1
BibTeX
@inproceedings{luss2012predicting,
  author = {Ronny Luss and Alexandre d’Aspremont},
  title = {{Predicting abnormal returns from news using text classification}},
  booktitle = {Quantitative},
  year = {2012},
  doi = {10.1080/14697688.2012.672762},
}
Figure