[2603.04403] FinRetrieval: A Benchmark for Financial Data Retrieval by AI Agents

Computer Science > Information Retrieval arXiv:2603.04403 (cs) [Submitted on 2 Jan 2026] Title:FinRetrieval: A Benchmark for Financial Data Retrieval by AI Agents Authors:Eric Y. Kim, Jie Huang View a PDF of the paper titled FinRetrieval: A Benchmark for Financial Data Retrieval by AI Agents, by Eric Y. Kim and Jie Huang View PDF Abstract:AI agents increasingly assist with financial research, yet no benchmark evaluates their ability to retrieve specific numeric values from structured databases. We introduce FinRetrieval, a benchmark of 500 financial retrieval questions with ground truth answers, agent responses from 14 configurations across three frontier providers (Anthropic, OpenAI, Google), and complete tool call execution traces. Our evaluation reveals that tool availability dominates performance: Claude Opus achieves 90.8% accuracy with structured data APIs but only 19.8% with web search alone--a 71 percentage point gap that exceeds other providers by 3-4x. We find that reasoning mode benefits vary inversely with base capability (+9.0pp for OpenAI vs +2.8pp for Claude), explained by differences in base-mode tool utilization rather than reasoning ability. Geographic performance gaps (5.6pp US advantage) stem from fiscal year naming conventions, not model limitations. We release the dataset, evaluation code, and tool traces to enable research on financial AI systems. Comments: Subjects: Information Retrieval (cs.IR); Artificial Intelligence (cs.AI); Computation and Lang...