Mergers and Acquisitions (M&A) remain a high-stakes strategic activity, yet empirical evidence suggests that approximately 70% to 90% of M&A deals fail to meet their initial strategic expectations. This study addresses the challenge of identifying early warning signs of deal failure by investigating the structural limits of pre-deal financial data in predicting post-M&A value creation, defined as six-month Buy-and-Hold Abnormal Returns (BHAR). Using a dataset of 463 U.S. public-to-public transactions (2010–2024), we benchmark linear models (Logistic Regression, SVM) against non-linear ensembles (XGBoost) under strict leakage controls. To mitigate the noise inherent in efficient markets, we introduce a Quantile Labeling Strategy, isolating the top and bottom 30% of outcomes. Our results document a predictive ceiling (AUC ≈ 0.60), supporting the Semi-Strong Efficient Market Hypothesis (EMH). However, feature importance analysis reveals a robust negative signal associated with Relative Deal Size, providing empirical support for the “Indigestion Hypothesis”—that integration complexity scales super-linearly with target size. Notably, simple linear models performed comparably to complex ensembles, suggesting that for low signal-to-noise financial datasets, model complexity yields diminishing returns.

Keywords

M&A, Machine Learning, Indigestion Hypothesis, XGBoost, Efficient Market Hypothesis.