Identifying M&A Targets from Textual Disclosures : A Transformer Neural Network Approach

Can textual information from firm disclosures help to identify M&A targets? We employ the state-of-the-art transformer neural network RoBERTa based on 130,000 annual financial reports of publicly listed US firms to estimate takeover likelihoods. We show that incorporating publicly available, highly standardized textual information can improve predictability of corporate takeovers significantly in out-of-sample tests, and that this information is not fully incorporated in prices on financial markets. Both Business Description and the Management Discussion and Analysis sections of Form 10-K filings contribute to the increased predictive performance. We use explainable artificial intelligence methods to examine the reasons for the improved predictions. Our analyses indicate that the model is able to identify firm-specific capabilities sought by acquirers as well as governance issues that lead to a change of corporate control. As a result, we are able to improve the identification of fundamentally motivated acquisitions, whereas mergers driven by speculative motives remain hard to predict