Global Merger-Arbitrage Forecasting with Language Models

We present a language-model forecasting system for merger arbitrage, a specialized high-stakes financial setting in which the task is to predict the outcome of announced M&A deals. Unlike prior work on judgmental forecasting with LLMs, which has focused on broad mixed-topic benchmarks and short context such as news snippets, we study a setting that requires long-context reasoning over hundreds of pages of technical documents. Our system combines expert-guided context engineering with finetuning on hindsight-guided reasoning traces derived from historical deals. Given an announced deal, it outputs a probability distribution over three mutually exclusive outcomes: closing at announced terms, a higher bid, or deal termination. On an out-of-sample set of more than 400 large deals spanning 42 countries, our finetuned system achieves the best performance of any method we evaluate, reducing class-balanced Brier score to 0.151. This is 24% below calibrated market-implied probabilities, 19% below XGBoost, and 25-42% below frontier language models. These results, together with ablation studies, show that LLM-based forecasting can succeed in specialized, long-context financial workflows, with hindsight-based supervision and expert-designed context playing a critical role.

Researchers have recently started using language models to forecast future events by reading text, rather than relying only on traditional models built from structured data. We test whether this ability could help with a real financial task: predicting the outcome of an announced merger or acquisition. When a merger is announced, investors must judge whether the deal will close as planned, be replaced by a higher offer, or fall apart. Accurate predictions can help them manage risk and make better decisions, but this is hard. A single deal may require reading long legal documents and combining information about regulators, how the deal is being paid for, other potential buyers, and broader market conditions. We built an automated language-model system for this task. First, we worked with experienced merger investors to design tools that research the most relevant information for each deal. Then, we then trained the model on past deals using what was known at the time and how each one ultimately ended. In tests on more than 400 deals the system had not seen before, it outperformed market-based signals, a traditional machine learning model, and leading general-purpose language models. Our work shows that AI forecasting can succeed in specialized, high-stakes settings when it is built around real expert workflows.