Predicting Shareholder Activism Risk: A Machine-Learning-Based Approach for Corporate Defense Strategy

Abstract

As shareholder activism continues to grow more prominent across the U.S. corporate landscape, the value of tools that can identify vulnerable firms in advance increases. This thesis studies whether machine-learning models, trained only on publicly observable firm characteristics, can predict such activism risk. In doing so, it develops a target-prediction framework for 7,768 U.S. publicly traded operating companies from 2010 to 2024. To identify the strongest framework, a model horse race across linear and nonlinear model families is performed using 73 predictors spanning market, accounting, valuation, activism-history, peer-relative, governance, and ownership blocks. Models were trained on 2010–2021 data and evaluated on a 2022–2024 hold-out. A secondary extension asks whether a slightly expanded set of 84 predictors could predict the outcome of campaigns conditional on launch.

In the primary horse race, all model families materially outperformed the naive prevalence benchmarks. A non-winsorized XGBoost specification ultimately emerged as the strongest overall model, attaining a holdout PR-AUC of 0.2865 and ROC-AUC of 0.7830. While recent activism history was the strongest predictive variable across both the linear and nonlinear winners, ownership, governance, peer-relative, and operating characteristics also contributed meaningfully to performance. Interpretability analysis further suggested that the XGBoost winner captured threshold effects and interactions beyond the leading linear benchmark. By contrast, the campaign outcome-prediction extension yielded weaker results, with the thesis unable to produce a strong or stable predictive framework. Taken together, these findings support the use of a machine-learning-based activism-risk screener for proactive corporate defense while also clarifying the limits of forecasting campaign resolution conditional on launch.