Source-linked AI summary
Predicting Abnormal Returns From News Using Text Classification
Ronny Luss, Alexandre d'Aspremont
TL;DR
The paper asks whether financial-news text can predict intraday abnormal returns more effectively than historical returns alone. It applies SVM text classification and multiple kernel learning to combine press-release word frequencies with equity returns, using an analytic center cutting plane solver. The results show that return direction is not predictable, but return size is, with text features producing strong performance and combined kernels improving classification.
Problem
Existing volatility models based on asset returns largely omit financial news, motivating a more refined analysis of news effects on asset prices.
Method
SVMs classify whether absolute returns jump after press releases using word frequencies and equity returns, while multiple kernel learning combines these feature sources.
Results
Multiple kernel learning greatly improved performance, while abnormal-return predictability was significant even though return direction was not predictable.
Takeaways & Limitations
Text features provide useful information about the size of intraday returns, and their predictive power decreases as the intraday horizon lengthens.
Takeaways & Limitations
The reported very high Sharpe ratios reflect a simple trading strategy and do not imply that equally high ratios can be generated in practice.
Abstract
from arXiv · showhide
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 Introduction
The paper frames financial news as an underused source for modeling market volatility and uses text classification to predict abnormal intraday returns. It combines press-release text with historical returns through SVMs and multiple kernel learning, solved with an analytic center cutting plane method.
- Asset-pricing jump processes and ARCH/GARCH volatility models do not finely incorporate financial news as a source of market volatility.The paper motivates text classification as a more refined way to analyze how news affects asset prices.
- Press releases are labeled by whether absolute returns jump after publication, while return direction is not found to be predictable.The classification horizons range from 10 to 250 minutes after a release.
- Support vector machines use word frequencies and equity returns to predict abnormal intraday returns, with predictability decreasing throughout the trading day.The study evaluates text and past absolute returns across many intraday horizons.
- Multiple kernel learning combines text and asset-price time-series features to enhance classification performance.The resulting optimization problem is solved with an analytic center cutting plane method.
- The paper extends prior text-classification work by analyzing abnormal-return predictability at many horizons rather than only short-window return categories.The comparison study used a 15-minute window and four return categories, whereas this paper examines horizons from 10 through 250 minutes.
2 Predictions with support vector machines
The paper uses SVMs to classify intraday abnormal returns from press-release text and past equity returns, finding abnormal-return predictability but little directional predictability. Text generally outperforms returns, while performance depends on the abnormal-return threshold, forecasting horizon, and trading assumptions.
- Method: SVMs classify whether an abnormal return will occur after a press release using either bag-of-words text features or past equity returns.Text uses a predetermined 619-word stemmed dictionary with term-frequency and document-frequency weighting.
- Directional returns: No predictability is found for return direction: Sharpe ratios remain near zero and accuracy stays close to 50% using either text or returns.The directional experiments use linear kernels with a single feature type.
- Abnormal returns: Abnormal returns are predictable from both absolute-return histories and press-release text, with text producing better predictions.Abnormal returns are defined using the 75th percentile of absolute returns observed in each training window.
- Time-of-day effects: Including early-day articles improves the ≥10:10 AM curve, which the authors attribute to models learning the time-of-day pattern in absolute returns.The text and returns experiments themselves do not use timestamp features, and day-of-week features perform weakly.
- Forecast horizon and threshold: Performance decreases as the intraday forecasting horizon increases because public information is absorbed into prices over time.The paper also reports strong sensitivity to the abnormal-return threshold: the 50th percentile sharply harms text performance, while the 85th percentile improves performance relative to the 75th.
- Trading implications: Trading results are constrained by strategy design and transaction costs: high reported Sharpe ratios may not be achievable in practice, and bid–ask spreads produced uniformly negative Sharpe ratios.The paper also reports asymmetric payoff risks between predicting abnormal returns and predicting no movement.
- Threshold sensitivity: 25–35% of press releases have news exerting a larger impact on future returns than past market data.This comparison concerns abnormal returns predicted 20 minutes ahead while varying the threshold percentile.
3 Combining text and returns
The paper combines text, returns, and time features through multiple kernel learning, using ACCPM to solve the kernel-weight optimization. The combined models improve abnormal-return prediction, while ACCPM offers stable computational behavior with fewer SVM computations in relevant settings.
- 3 Combining text and returns: MKL optimally combines text, return, and time-stamp kernels for abnormal-return classification.The candidate kernel set includes linear and Gaussian text and absolute-return kernels, plus time-of-day, day-of-week, and identity kernels.
- 3.2 Multiple kernel learning via an analytic center cutting plane method: ACCPM iteratively shrinks a localizing set around the optimum using first-order cuts and does not require differentiability.The method computes analytic centers and retains linear convergence while requiring fewer SVM computations than gradient-based approaches.
- 3.3 Computational Savings: ACCPM is more efficient than simpleMKL on the text-classification example because warm-starting does not offset simpleMKL’s additional SVM computations.For accpmMKL, the number of SVMs equals the number of iterations, whereas simpleMKL performs many SVMs during line searches.
- 3.3 Computational Savings: SimpleMKL can outperform ACCPM when the classification task is extremely easy and the optimal kernel mix is a singleton.ACCPM remains faster with LIBSVM in that example and is more stable across data sets and precision targets.
- 3.4 Predicting abnormal returns with text and returns: Combining text, absolute returns, and timestamps significantly improves both accuracy and annualized daily Sharpe ratio over text-only or returns-only predictions.The evaluation uses the 75th percentile of training absolute returns as the abnormality threshold.
- 3.4 Predicting abnormal returns with text and returns: Across 50th and 85th percentile thresholds, multiple kernels slightly improve performance and retain a mixture of text and absolute-return kernels.Equally weighted or highly suboptimal kernels reduce performance, while random kernels have negligible coefficients.
4 Conclusion
The paper finds that combining text with absolute returns significantly predicts abnormal returns, while multiple kernel learning improves performance. It identifies unavailable intraday options data and simple word selection as boundaries, and proposes regression analysis as a natural extension.
- Significant performance was achieved in predicting abnormal returns using text and absolute returns, with multiple kernel learning greatly improving performance.The paper also describes a cutting-plane algorithm for large-scale multiple kernel learning and demonstrates its efficiency relative to current solvers.
- Intraday options prices were unavailable, preventing realistic performance metrics from a tradeable strategy based on abnormal-return predictions.The authors note that equity options would provide realistic metrics, but their intraday prices are not publicly available.
- Feature selection remains an important research direction because the experiments used a simple handpicked set of dictionary words.More advanced semantic methods and industry-specific dictionaries are proposed for future work.
- Support vector regression with multiple kernel learning could extend the approach to forecasting both intraday volatility and abnormal returns.