FinDeepForecast introduces the first live, end-to-end multi-agent system for benchmarking deep research (DR) agents in financial forecasting. The system features a dual-track taxonomy that dynamically generates both recurrent (e.g., quarterly earnings) and non-recurrent (e.g., M&A impact) forecasting tasks at corporate and macro levels. Using this system, the authors create FinDeepForecastBench, a weekly evaluation benchmark spanning ten weeks, covering 8 global economies and 1,314 listed companies. They evaluate 13 representative methods, including state-of-the-art DR agents and traditional baselines.

Extensive experiments reveal that DR agents consistently outperform strong baselines, demonstrating their potential in research-oriented financial tasks. However, their performance still falls short of genuine forward-looking financial reasoning, indicating room for improvement. The benchmark and leaderboard are publicly available on the OpenFinArena platform, providing a standardized testbed for future advancements in financial forecasting with AI agents.