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.
Introduction
Financial markets are fundamentally driven by information. While traditional volatility models like ARCH and GARCH rely heavily on historical price data, they often overlook the qualitative "shocks" introduced by news releases. This research investigates whether intraday "abnormal returns"—price movements significantly larger than typical market noise—can be predicted by integrating news text with structured time-series data.
Methodology: SVM and Multiple Kernel Learning
The core of the approach involves framing the prediction task as a binary classification problem. The researchers utilize Support Vector Machines (SVM) to determine if a press release will trigger an abnormal return within a specific time window.
To fuse disparate data types, the study employs Multiple Kernel Learning (MKL). This allows the model to automatically learn the optimal combination of various "kernels" or similarity functions representing text (TF-IDF bag-of-words), historical returns, and temporal features like time-of-day. A significant technical contribution is the introduction of the Analytic Center Cutting Plane Method (ACCPM) for MKL optimization. ACCPM iteratively refines a localization set to find the optimal solution, proving far more efficient than standard gradient-descent solvers like SimpleMKL for high-dimensional financial datasets.
Data and Experimental Design
The study leverages a robust dataset comprising eight years of press releases from PRNewswire (2000–2007) and intraday stock prices from the NYSE Trade and Quote (TAQ) database.
The researchers tested prediction horizons ranging from 10 to 250 minutes. As shown in the data distribution, the volume of available training and testing data varies across these windows, providing a comprehensive view of how news impact decays or manifests over time.
Key Insights and Results
The findings challenge and refine aspects of the Efficient Market Hypothesis:
- Magnitude vs. Direction: Predicting the direction of a stock move (up vs. down) remains near a 50% "coin flip." However, the magnitude of the move—predicting that a significant jump will occur—is highly achievable.
- Text as a Superior Signal: Textual features from press releases are significantly stronger predictors of future volatility than historical price data alone.
- The Time-of-Day Effect: Market volatility naturally spikes at opening and closing. By including a specific "time-of-day" kernel, the MKL model successfully distinguishes between routine volatility and news-driven abnormal returns.
- Computational Efficiency: The ACCPM algorithm achieved substantial speedups, often performing orders of magnitude faster than state-of-the-art solvers, making it viable for large-scale financial applications.
Conclusion
This research demonstrates that combining unstructured news text with structured financial data through MKL and efficient optimization significantly enhances the predictability of market volatility. While direction remains elusive, the ability to forecast abnormal returns offers substantial value for risk management and algorithmic trading. Future work may explore more advanced text modeling techniques like Latent Dirichlet Allocation (LDA) or extending the classification framework to regression for precise magnitude estimation.
Summary updated