LLM system beats market in merger arbitrage forecasting by 24%
New finetuned model outperforms XGBoost and frontier LLMs on 400+ global deals
A team of researchers (Jajal et al.) has introduced a novel LLM-based forecasting system for merger arbitrage, a high-stakes financial task requiring prediction of M&A deal outcomes. Unlike prior work on broad mixed-topic benchmarks, this system tackles long-context reasoning over hundreds of pages of technical documents (e.g., merger agreements, regulatory filings). The model combines expert-guided context engineering with fine-tuning on hindsight-guided reasoning traces from historical deals. It outputs a probability distribution over three mutually exclusive outcomes: closing at announced terms, a higher bid, or deal termination.
Tested on an out-of-sample set of over 400 large deals spanning 42 countries, the finetuned system achieves a class-balanced Brier score of 0.151—24% below calibrated market-implied probabilities, 19% below XGBoost, and 25–42% below frontier language models (e.g., GPT-4). Ablation studies confirm the critical role of hindsight-based supervision and expert-designed context. Accepted to ICML 2026, this work demonstrates that LLMs can succeed in specialized, long-context financial workflows, potentially transforming how hedge funds and investment banks approach merger-arbitrage strategies.
- System predicts three deal outcomes: closing, higher bid, or termination with a Brier score of 0.151
- Outperforms market-implied probabilities by 24% and XGBoost by 19% on 400+ large deals
- Uses expert-guided context engineering and hindsight reasoning traces over hundreds of pages of documents
Why It Matters
LLMs can now outperform traditional financial models in complex, long-context forecasting tasks like merger arbitrage.