言語モデルによるグローバルM&Aアービトラージ予測Global Merger-Arbitrage Forecasting with Language Models

A new paper posted to arXiv examines whether large language models (LLMs) can improve the prediction of merger-and-acquisition (M&A) deal outcomes, a task at the heart of the merger-arbitrage investment strategy. The work argues that natural language processing (NLP) signals extracted from financial text can meaningfully raise forecasting accuracy across global markets, suggesting a broader role for language models in quantitative finance.

Merger arbitrage is built around a single uncertain event: whether an announced acquisition will actually close. When a company announces plans to buy another, the target's share price typically jumps but usually settles below the offered price. That gap, known as the arbitrage spread, reflects the market's collective judgment about the risk that a deal collapses, is renegotiated, or is delayed. Arbitrageurs buy the target's shares to capture the spread if the transaction completes, and their returns depend heavily on correctly estimating the probability of completion. Misjudging a deal that later breaks can produce sharp losses, so accurate outcome prediction is central to the approach.

Historically, models for deal completion have relied on structured variables such as the size of the premium offered, the method of payment (cash versus stock), whether the bid is hostile or friendly, the regulatory jurisdiction, and the presence of competing bidders. The paper's contribution is to test how much additional signal can be drawn from unstructured text, including regulatory filings, merger agreements, press releases, and news coverage. Because these documents contain nuanced language about financing conditions, antitrust exposure, shareholder approval, and management intent, they are a natural target for models designed to interpret human language.

According to the summary, the authors find that features derived from LLMs improve predictive performance relative to approaches that ignore textual information. The framing across global markets indicates the study looks beyond a single country, which matters because deal completion dynamics differ by jurisdiction. Regulatory regimes, disclosure requirements, and the prevalence of government intervention vary widely between, for example, US, European, and Asian markets, and a model that generalizes across them would be more useful than one tuned to a single exchange.

The approach fits within a wider trend of applying language models to finance. Domain-specific systems such as FinBERT, trained to classify sentiment in financial text, and larger efforts like BloombergGPT, built on a broad corpus of financial documents, reflect sustained interest in adapting NLP to market data. Hedge funds and quantitative asset managers have for years mined news feeds, earnings-call transcripts, and filings for tradable signals, and the arrival of general-purpose LLMs has lowered the barrier to extracting structured insight from messy text. This paper appears to extend that lineage to the specific, event-driven problem of merger arbitrage.

Several caveats are worth keeping in mind when interpreting results of this kind. Backtested improvements in prediction accuracy do not automatically translate into trading profits, because arbitrage spreads already embody much of the market's information and because transaction costs, position limits, and the timing of deal breaks can erode theoretical gains. There is also the risk of look-ahead bias if a model is exposed, even indirectly, to information that would not have been available at the time a decision was made, a particular concern with LLMs whose training data may overlap with the historical period being tested. Robust evaluation typically requires careful separation of training and test windows and attention to how signals decay over time.

The work also touches on a broader methodological question: how to convert an LLM's contextual understanding into features that a forecasting pipeline can use reliably. Options generally include using model-generated embeddings, prompting a model to produce classifications or probability estimates, or fine-tuning on labeled deal outcomes. Each carries trade-offs in interpretability, cost, and stability. For practitioners, transparency about why a model expects a deal to fail can matter as much as the prediction itself, especially in a regulated investment setting.

As a research contribution, the paper adds to evidence that language models can surface useful information from financial documents that structured data alone may miss. Whether the reported gains hold up out of sample, across market cycles, and net of real-world frictions will likely determine how much the approach influences practice. For now, it is best read as a demonstration of potential rather than a finished trading system.