Title: Hedge-Bench: Evaluating Agents on Complex, Real-World Financial Reasoning Tasks

Abstract:

While AI agents are becoming increasingly proficient in the mechanical aspects of financial analysis—such as document retrieval, formula calculation, and spreadsheet updates—the more demanding and high-value challenge remains reasoning through open-ended questions that characterize the work of expert analysts. Current benchmarks fail to adequately represent this specific category of problems. Furthermore, existing evaluations of open-ended reasoning often depend on model-judged outputs, which introduce noise and circularity into the assessment process.

To address these gaps, we introduce Hedge-Bench 1.0. This benchmark comprises 102 authentic, on-the-job tasks derived from the explicit reasoning traces of professional hedge fund analysts who utilized relevant information sources. This methodology allows for deterministic grading based on verified expert steps. Our results show that frontier models and agents achieve scores below 16% on this benchmark. The dataset and evaluation harness are publicly available at github.com/Trata-Inc/trata-hedge-bench.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC