Frontier Financial Judgement: Can agents tell what might move a stock?

As the volume of financial information grows, equity analysts increasingly rely on AI to filter news and identify market-moving events. This paper investigates whether current AI agents possess the financial "taste" required to distinguish genuinely new, valuation-relevant information from stale, misleading, or immaterial news. The authors seek to determine if these agents can replicate the nuanced judgments of professional equity analysts under realistic, noisy conditions.

The authors introduced "Frontier Financial Judgement," a new benchmark developed with professional analysts. The dataset consists of 82 expert-designed, realistic synthetic events mixed with live news and historical documents to create 656 assessment items. This setup forces agents to perform complex tasks—such as identifying recycled disclosures, interpreting segment-level margin guidance, and separating factual signals from adversarial framing—while using web search tools to verify information against a fixed, point-in-time cutoff. The researchers evaluated 14 different agents, measuring their accuracy across three dimensions: information novelty, expected valuation importance, and directional impact.

The study reveals a substantial gap between current AI capabilities and the requirements for reliable financial news filtering. Even the most advanced agents failed to match expert labels in nearly half of all cases, with the top performer achieving only 52.4% all-label accuracy. Furthermore, the researchers observed a wide divergence in false-positive rates, ranging from ~1% to ~32%. This indicates that high accuracy on synthetic targets does not guarantee reliability in real-world environments, where agents must filter out high volumes of irrelevant "noise" without flagging stale information as significant.

Reliable news-flow filtering is essential for modern equity analysis, yet this research demonstrates that current models struggle with the fundamental financial reasoning required to process market data. The significant trade-offs between accuracy, cost, and reliability suggest that automated news-filtering systems are not yet ready for deployment in professional financial workflows. The benchmark provides a necessary, rigorous standard for future development in financial AI.