Predicting merger withdrawals is challenging because failed deals can impose large costs on firms and investors, while withdrawals are rare and difficult to identify using structured deal variables alone. We propose MergerLLM, a large-language-model-based framework that predicts merger completion using serialized deal characteristics and text-based firm information. Using a sample of U.S. mergers and acquisitions (M&A) deals, we compare MergerLLM with machine-learning and deep-learning benchmarks. MergerLLM improves detection of withdrawn deals, especially in recall and F1-score, while maintaining competitive ranking performance. Its predicted merger-success probabilities also contain economic information in post-announcement return tests. We then construct HAIDiff, a proxy for model–market disagreement, defined as the difference between MergerLLM’s predicted merger-success probability and a transformed announcement-return-based market signal. Climate exposure is positively associated with this disagreement measure, and the result is robust to an alternative measure based on target cumulative abnormal returns. The evidence suggests that climate-exposed deals are settings in which model-implied completion probabilities and market-based signals diverge more strongly.
Shohrab et al. (Wed,) studied this question.