Global Merger-Arbitrage Forecasting with Language Models
A ReAct retrieval and fine-tuned ensemble reaches 0.151 class-balanced Brier on 404 M&A deals, outperforming markets, XGBoost, and frontier LLMs.
Key Findings
Methodology
The system uses a two-stage design: 12 GPT-4.1 or GPT-5 ReAct research agents retrieve and analyze filings, ownership, financing, regulation, market commentary, and precedent deals; a fine-tuned frontier LLM then synthesizes 6.8k–10k tokens of structured context. Hindsight-guided reasoning traces are generated from realized historical outcomes. Five samples at temperature 0.2 are aggregated by the median to produce three outcome probabilities, completion time, and a cited report.
Key Results
- On 404 out-of-sample deals announced from February to December 2025 across 42 countries, the system achieved a class-balanced Brier score of 0.151. This compares with 0.199 for Platt-scaled market-implied probabilities, 0.186 for XGBoost, and 0.201–0.259 for frontier models, corresponding to improvements of about 24%, 19%, and 25–42%.
- Fine-tuning reduced calibration error from 0.089 to 0.039. Forecast correlation with market-implied probabilities was 0.36, versus 0.75 for XGBoost, indicating that the system was not merely reproducing market consensus.
- Ablations increased BrierB by 0.032 without hindsight guidance, 0.021 when restricted to Deal Card information, and 0.026 when half of the 1,244-deal training set was used. Supervision, context breadth, and sample size were therefore all material.
Significance
The paper moves LLM judgmental forecasting from short news snippets and heterogeneous question banks into a specialized, high-stakes, long-document financial workflow. For industry, the system can support position sizing, risk management, and portfolio construction; for research, it shows that expert-designed retrieval and outcome-informed supervision can add information beyond structured statistical baselines. The authors deploy it as decision support, not autonomous trading.
Technical Contribution
The technical contributions are a domain-specific 12-agent ReAct pipeline, a million-document corpus with fine-tuned embeddings and passage reranking, strict forecast-date retrieval constraints, and fine-tuning targets derived from post-mortem deal analysis. The evaluation combines three-way deal outcomes, market-implied probabilities, Murphy decomposition, and class-, surprise-, and P&L-weighted Brier scores. This creates a practical framework that evaluates calibration, ranking, unusual outcomes, and economic relevance rather than raw accuracy alone.
Novelty
Unlike Halawi et al.’s news-retrieval forecasters, AIA Forecaster’s general search-and-ensemble system, and structured XGBoost baselines, this work integrates specialist merger-arbitrage workflows, long regulatory documents, and hindsight-guided supervision. Its novelty is primarily systems-level rather than a new foundation model: it demonstrates that expert context engineering plus temporally safe fine-tuning can deliver measurable gains in a professional long-context forecasting task.
Limitations
- The dataset contains 1,648 large public-target deals from only 2022–2025, with severe imbalance and just about 3% Fail+ outcomes. Results may therefore depend on the sampled macroeconomic, geographic, and regulatory regimes and may not generalize to small or older transactions.
- The market-implied baseline uses a simplified two-state valuation with estimated upside and failure values. Actual break prices depend on termination causes, alternative bids, and market conditions, so the baseline is informative but structurally imperfect.
- The architecture requires 12 agents, commercial data, long contexts, and five forecast samples. Latency, token cost, data licensing, and expert maintenance may limit real-time or broad deployment.
Future Work
Future work should expand the historical horizon and transaction universe, report performance by country, sector, and regulatory regime, and compare online updating, calibration methods, and cost-constrained agent configurations. Prospective live evaluation is especially important: researchers should test whether reports improve position sizing, risk-adjusted returns, and portfolio outcomes rather than only offline Brier scores. Robustness under regime shifts and rare outcomes also remains central.
AI Executive Summary
Merger arbitrage asks a deceptively precise question: after an acquisition is announced, will it close on the stated terms, be replaced by a higher bid, or collapse with losses for target shareholders? Answering it requires far more than reading a headline. Analysts may need to inspect a hundred-page merger agreement, regulatory filings across jurisdictions, shareholder voting dynamics, financing capacity, litigation, and comparable transactions. Existing LLM forecasting studies have largely used short news context and mixed-topic benchmarks, leaving their value in specialist, long-context finance uncertain.
Jajal and colleagues build a forecasting system around the actual workflow of merger-arbitrage professionals. Twelve GPT-4.1 or GPT-5 ReAct agents investigate distinct dimensions, including filings, ownership, regulatory risk, financing, precedent mergers, market views, and future catalysts. Their outputs are assembled into a cited context of typically 6.8k–10k tokens. A fine-tuned forecasting model then begins with historical base rates, weighs deal-specific red and green flags, and predicts three mutually exclusive outcomes, completion time, and a report. Training uses hindsight-guided traces generated from historical post-mortems, while timestamped retrieval and knowledge cutoffs prevent test-period leakage.
The results are unusually strong for this specialized setting. On 404 held-out deals spanning 42 countries, the best model reached a class-balanced Brier score of 0.151, versus 0.199 for Platt-scaled market probabilities, 0.186 for XGBoost, and 0.201–0.259 for frontier LLMs given identical context. Fine-tuning reduced calibration error from 0.089 to 0.039, while forecast correlation with market probabilities remained only 0.36, compared with 0.75 for XGBoost. Ablations show that hindsight supervision, broad context, and training scale each matter. The system is therefore best understood as a high-speed, auditable research assistant—not a self-governing trader. Its short historical window, rare-event imbalance, data costs, and offline evaluation leave important questions for prospective deployment.
Deep Analysis
Background
Judgmental forecasting research has shown that LLMs can approach non-expert human performance when supplied with retrieved news. Representative work includes Halawi et al. (2024), ReAct-based forecasting by Hsieh et al. (2024), Schoenegger et al. (2024, 2025), and AIA Forecaster by Alur et al. (2025). Yet these studies generally combine sports, politics, climate, macroeconomics, and other topics, often using shallow context. This paper instead treats merger arbitrage as a coherent specialist domain requiring technical documents, structured tools, and expert workflows.
Core Problem
The system predicts three mutually exclusive outcomes: Succeed+—closing as announced; Fail+—termination followed by a higher bid; and Fail−—termination without a favorable alternative. It also predicts days to completion and produces a cited report. The problem is difficult because Succeed+ represents 84.0% of the sample, Fail+ only 2.7%, and deal outcomes are idiosyncratic. Relevant evidence is distributed across long contracts, filings, shareholder records, regulation, financing, litigation, and changing market conditions.
Innovation
- �� A 12-agent ReAct workflow separates evidence collection from probabilistic synthesis. • A domain-fine-tuned embedding model and reranker search millions of timestamped documents rather than generic web snippets. • Hindsight-guided post-mortems identify which evidence deserved weight given the realized outcome, creating fine-tuning targets. • Date-restricted retrieval, pre-test knowledge cutoffs, and exclusion of open-web feeds reduce leakage. • Five low-temperature samples are median-aggregated, and GPT-5 cross-checks reports against source documents before finalization.
Methodology
- �� Dataset: 1,648 large public-target deals from 2022–2025, split temporally into 848 training, 396 validation, and 404 test deals. • Sources: definitive merger agreements, proxy statements, S-4 and 8-K filings, earnings transcripts, shareholder registries, historical M&A data, security mappings, and research commentary. • Agents: Ticker Resolution and Deal Card structure entities and terms; Filings, Ownership, Market View, Current Climate, Regulatory Risk, Financing Risk, Precedent Mergers, and Timeline Events analyze evidence; Gap Analysis finds omissions; Catalyst Tracker identifies forward risks. • Forecasting: a median 8.3k-token context is passed to the fine-tuned LLM, which outputs three probabilities, completion days, and a report. • Evaluation: Brier, class-balanced BrierB, surprise-weighted BrierS, P&L-weighted Brier$, Murphy decomposition, market alignment, and MAPE. Market-implied probability is inferred from target price, upside value, and beta-adjusted failure value.
Experiments
Deals are split strictly by announcement date: training ends January 2024, validation ends January 2025, and testing covers February–December 2025, producing 1,115 test forecast instances. Baselines include Platt-scaled market-implied probabilities, a rich-feature XGBoost model, and frontier LLMs such as GPT-5 under identical context. The test set spans 42 countries and has outcome frequencies of 84.0% Succeed+, 3.0% Fail+, and 14.4% Fail−. Ablations remove hindsight guidance, restrict context to Deal Card information, or halve the 1,244-deal training set.
Results
The best model achieved BrierB=0.151, compared with 0.199 for market probabilities, 0.186 for XGBoost, and 0.201–0.259 for frontier LLMs. Improvements were approximately 24%, 19%, and 25–42%, respectively. Calibration error fell from 0.089 to 0.039 after fine-tuning. Market correlation was 0.36 for the proposed system versus 0.75 for XGBoost. Removing hindsight guidance, broad context, or half the training data worsened BrierB by 0.032, 0.021, and 0.026, confirming that each component contributes materially.
Applications
Merger-arbitrage funds can use the system to screen announced transactions, size positions, identify regulatory and financing risks, and monitor voting, court, approval, financing, and competing-bid catalysts. Cited reports can accelerate analyst review of termination fees, lawsuits, ownership concentration, and deal-protection provisions. Deployment requires licensed timestamped data, robust entity resolution, human verification, and compliance controls. The paper’s practical positioning is decision support for analysts and portfolio managers, not unattended trade execution.
Limitations & Outlook
The four-year dataset is concentrated in large public-target transactions and contains very few Fail+ cases, limiting rare-event confidence and generalization across regimes. The market baseline simplifies outcomes into upside and downside states even though break prices depend on termination reasons and alternative bids. The 12-agent architecture, long contexts, five-sample ensemble, commercial corpus, and fact-checking stage create substantial token, latency, and maintenance costs. Excluding open-web content reduces leakage but may omit information used by real traders. Prospective, cross-cycle, cost-aware evaluation is still needed.
Plain Language Accessible to non-experts
Imagine a restaurant whose job is to decide whether a complicated catering order will actually be delivered. The order has been announced, but delivery could happen as promised, another customer could offer more money and take it over, or the order could be cancelled and leave the original customer worse off.
Instead of asking one hurried cook to guess, this restaurant employs twelve specialists. One reads the contract, another checks whether the rules allow the delivery, another asks who supports it, another checks whether the buyer can pay, and another studies similar orders from the past. They use only information that was available at the time, so they cannot peek at the ending.
A head chef combines the reports. The chef first remembers how often similar orders succeeded, then balances helpful clues against warning signs. The system makes five slightly different judgments and uses the middle one, reducing the influence of an odd answer. It returns three chances, an expected completion time, and an explanation with sources.
The researchers trained it by reviewing old orders after their outcomes were known and marking which clues mattered most. On 404 new deals from 42 countries, its error score was 0.151, compared with 0.199 for market-based estimates and 0.186 for another machine-learning method. It is not a crystal ball: it needs expensive information and human checking. Its real value is acting like a fast, organized research team that helps people make better decisions.
ELI14 Explained like you're 14
Think of a huge school trade: one school announces that it will buy another school. Everyone wants to know what happens next. Will the deal finish? Will a richer school jump in with a better offer? Or will the deal fail and make the smaller school’s situation worse?
A normal chatbot might read one headline and guess. This paper builds twelve detective bots instead. One checks the contract, one studies whether teachers and students—like shareholders—support the deal, one investigates whether regulators might block it, one checks whether the buyer has enough money, and one compares older school trades. They then send their notes to a final “coach” bot.
The coach gives three percentages, predicts how many days the deal may take, and explains its reasoning. It also studies old deals after knowing what happened, like watching game replays to learn which moves were important. But during the real test it is not allowed to look into the future. Five guesses are made and the middle guess is kept, so one weird answer does not dominate.
The results are impressive: on 404 deals from 42 countries, the system scored 0.151, while market estimates scored 0.199. Lower is better. Still, it is not a fortune teller! Deals can change suddenly, documents can be incomplete, and the system needs costly information. Think of it as a super-fast teammate who organizes evidence and spots danger—not someone who should place bets alone.
Glossary
Merger arbitrage
An investment strategy that buys a target after an acquisition announcement to capture the gap between its current price and expected deal value. The main danger is a failed transaction and a sharp price decline.
The financial setting and economic motivation for the forecasting task.
ReAct
A tool-using agent pattern in which a language model alternates between reasoning and actions such as search or structured lookup. It lets the model gather evidence before producing an answer.
Used by all 12 specialized research agents, following Yao et al. (2022).
Class-balanced Brier score
A proper scoring rule based on squared probability error, adjusted so rare and common classes contribute equally. Lower values indicate better probabilistic accuracy and balance.
The primary reported metric; the best value is 0.151.
Market-implied probability
A probability inferred from the target share price, estimated upside value, and expected value after failure. It represents the market’s aggregate assessment, not a guaranteed truth.
Used as both a baseline and an input or training signal.
Hindsight-guided supervision
Training supervision created by looking at realized historical outcomes and identifying which evidence should have received the greatest weight. It is applied to training deals only, not test deals.
The central fine-tuning strategy for producing useful reasoning traces.
Temporal integrity
The requirement that a forecast use only information available by its forecast date, including model knowledge and retrieved documents. It prevents future facts from contaminating evaluation.
Implemented with timestamped retrieval, knowledge cutoffs before testing, and exclusion of open-web feeds.
Open Questions Unanswered questions from this research
- 1 The model has not been shown to remain stable across longer histories, smaller transactions, different interest-rate regimes, or private deals. Cross-cycle and prospective evaluations are needed.
- 2 Fail+ represents only about 3% of observations. Reliable calibration for such rare alternative-bid events remains unresolved and may require specialized sampling or hierarchical models.
- 3 Offline Brier improvements do not prove better portfolio outcomes. Live or randomized deployment is needed to test effects on sizing, drawdowns, risk-adjusted returns, and analyst behavior.
Applications
Immediate Applications
Merger-arbitrage research assistant
Funds can use the system to organize merger agreements, regulatory requirements, ownership, financing, and precedent deals, then review a cited three-outcome forecast. It can prioritize research and expose missing evidence before a human decides whether to trade.
Risk and catalyst monitoring
Portfolio managers can use Timeline Events and Catalyst Tracker to monitor votes, approvals, court dates, financing deadlines, and competing-bid windows. New evidence can trigger human reassessment of probabilities, position size, and downside exposure.
Long-term Vision
Specialized financial forecasting platform
The architecture could extend to bankruptcy, activism, special situations, and regulatory-event forecasting. Major obstacles include licensed data, model drift, rare outcomes, computational cost, auditability, and the need for real-time compliance and human accountability.
Abstract
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.