What is the Deal? Predicting M&A outcomes with machine learning

ABSTRACT

We examine whether machine learning algorithms that incorporate accounting fundamentals, deal characteristics, and macro-economic indicators can predict which deals will be value creative versus value destructive, and offer three main results. First, non-linear machine learning models predict two-year post deal announcement returns relatively well. A trading strategy that buys the stock of the deals with the highest predicted score quintile and sells short the stock of the deals with the lowest predicted score quintile generates market adjusted returns around 11.9%. Second, the link between the machine learning prediction score and post-acquisition accounting earnings and cash flows is not direct. This suggests that the deals the market views as successful are not directly linked to earnings or cash flows in the immediate three years after the deal closes. Finally, we find that the one-year return spread between the best and worst performing deals is decreasing over time. These results suggest that market participants are becoming more sophisticated with their acquisition choices, perhaps as they are able to use machine learning themselves. Overall, we shed light on the M&A outcome puzzle that has persisted in the literature for decades, suggesting that one reason these papers have been unable to clearly predict future M&A success is due to being limited to linear models and in some cases due to linking deal success to accounting earnings or cash flows in the first three years post-acquisition.