This study applies machine learning models such as TabNet, XGBoost, CatBoost, Support Vector Machines, a Multilayer Perceptron, and Logistic Regression to predict takeover-rumor accuracy using a proprietary dataset of 2,074 rumor articles with identifiable target firms, screened from over 30,000 news articles from January 2002 to December 2011. In the raw feature (no-PCA) specification, Logistic Regression and TabNet perform similarly. After addressing class imbalance and applying dimensionality reduction (PCA), TabNet materially outperforms Logistic Regression and delivers economically meaningful gains: A TabNet-based long-short strategy earns higher average monthly abnormal returns than a Logistic Regression strategy at both horizons, 1.148% versus 0.773% with a six-month holding period and 1.074% versus 0.841% with a one-year holding period. SHAP analyses show that abnormal returns around the rumor date and the informativeness of rumor content are the most influential predictors, consistent with market reactions reflecting real-time credibility assessments and informative narratives providing more verifiable signals of deal realization. Partial dependence patterns reveal non-linear relationships for key features, highlighting limits of purely linear specifications. Overall, the results show how interpretable machine learning can convert noisy information events into transparent credibility scores that improve prediction and decision-making in financial markets.
Khadivar et al. (Wed,) studied this question.