FinSkillBench Evaluates AI Agents for Investment Management Tasks

A new evaluation suite, FinSkillBench, has been introduced to assess the proficiency of language model agents in performing complex investment management tasks. Unlike general text generation, this domain demands precise data retrieval, accurate computational inputs, specialized method invocation, and auditable structured outputs. The benchmark covers three critical financial domains: portfolio construction, risk management, and fundamental analysis, encompassing 12 subtasks and 2,603 task episodes. Each episode provides point-in-time inputs, a hidden ground truth, and a task-specific verifier to ensure rigorous evaluation. Evaluations across nine models demonstrated that agents equipped with curated skill packages, consisting of procedural documents and executable components, consistently outperformed those without skills, significantly raising mean scores. Conversely, agents attempting to self-generate skills showed minimal improvement despite higher computational costs, suggesting that reliable, pre-defined procedural skills are crucial for effective AI application in high-stakes financial contexts.