We introduce Investor's Last Exam (ILE), an equity‑research benchmark designed with input from seasoned buy‑side investors and fund managers. ILE consists of three main components:
-
Versioned financial database containing 3.43 million document versions across nine source categories, supplemented by structured market data. Every entry carries a validated timestamp to strictly enforce information‑access cut‑offs for retrieval and computation.
-
Task taxonomy organized around professional analyst deliverables—company overviews, financial models, investment recommendations—resulting in 86 concrete assignments.
-
Expert‑led scoring framework that links evaluation criteria, data sources, calculations, and assumptions through financial dependency graphs, enabling joint assessment of research quality, forecast accuracy, and simulated investment performance.
In the evaluation, GPT‑6 Astra and Claude Opus 5.5 achieved overall scores of 68.1 and 66.9 out of 100, yet their "Forecast & Model" sub‑scores hovered around 52.5, indicating that even state‑of‑the‑art models struggle to translate comprehensive research capability into reliable predictions.
ILE shifts the evaluation focus from isolated capabilities to the evidence‑assumption‑judgment chain, allowing researchers to pinpoint where analytical strengths fail to produce actionable forecasts and providing a traceable, diagnostic foundation for training and assessing financial agents.
Review