A Comparative Study of Logistic Regression and Machine Learning to Identify Acquirer Success Factors

Abstract

This paper develops, presents and tests two research questions that contribute to current explanations of shareholder wealth creation in mergers and acquisitions transactions. We (l) identify pre-acquisition success factors and (2) evaluate their practical usefulness for managers seeking to acquire other firms. We build on Cartwright and Schoenberg (2006)'s framework for understanding the persistent failings of acquisitions. This includes agency problems, research not reaching practitioners and the need for new methods to explain M&A success. Our findings indicate that financial ratios play a significant role in determining the success of acquirers. We develop and validate both a logistic regression and two machine learning models, revealing significant factors that impact acqmrer success. Our results from the logistic regression mirror those from much of existing literature, identifying several significant factors for acquirer success. Furthermore, we find support for the prevalence of agency problems in acquisition decisions (Jensen, 1986; Maloney et al., 1993) and the internal market hypothesis (Stein, 1997; Shin and Stulz, 1998). Yet, our results also conflict with existing literature on several points. While our logistic regression reveals statistically significant acquirer success factors, its poor predictive performance makes it impractical for managers in real-world applications. In contrast, our machine learning methods identify complex non-linear relationships and discriminates well between successful and unsuccessful acquirers, resulting in ROC curves with excellent AUC scores. This supports the argument that the true relationships between acquirer success and the predictors are too complex for a logistic regression approach, even though much of existing literature on the subject builds on the logistic regression. We thus provide a possible explanation for why M&A success rates are still low, despite the extensive research on the subject. Finally, we argue that the key to enabling managers to use machine learning models directly lies in the adoption of partial dependence plots, as they facilitate a deeper understanding of the models and lets managers explain them to stakeholders more easily.