Global Merger-Arbitrage Forecasting with Language Models
This study addresses the challenge of predicting the ultimate outcomes—completion under original terms, acquisition by a higher bidder, or termination—of announced merger and acquisition (M&A) deals, which requires processing long-context information from hundreds of pages of technical documentation. The authors propose a novel approach that integrates expert-guided context engineering with fine-tuning based on reasoning trajectories derived from post-hoc analyses of historical transactions, complemented by probabilistic calibration and a multi-class evaluation framework. This method represents the first successful application of large language models to M&A outcome prediction in a highly specialized, long-context financial setting, achieving a class-balanced Brier score of 0.151 across more than 400 cross-border, large-scale transactions—significantly outperforming market-implied probabilities, XGBoost, and current state-of-the-art language models.