Investor's Last Exam: Benchmarking AI Equity Research Without Future Leakage

Jiang, Hu, Sheng, Cai, Ge, Kong, Lin, Chen, Weng, Yan · Prophetlab / RethinkLab, SJTU, USTC, Foresight Fund, IvyRock · Oct 2026
arXiv 2610.38342

The Tesla leak: date filter vs version-level availability

Task O74, t0 = March 24, 2026. Reported: a "before March 24" search returned a July 2 page with realized Q2 deliveries of 480,126. Positions of the first version are illustrative.Illustrative layout

Frozen world: every tool call passes one gate

Click a tool or step through. Search, read, download, table tools and cached results all use the same check.

Database scale

Reported figures: about 3.4M versions, 1.52 TB, nine source categories. Per-category split not given.

Dependency graph: refinancing example

Cash + operating cash flow + unused credit looks sufficient, until a covenant makes the credit undrawable.

Three yardsticks

The paper's claim: they do not rank models the same way.

Interval scoring fix (AWS O15)

Reported: narrow miss earned 15.8 points; revised 80% central interval, joint accuracy and width, scored 0 on a miss.

Results explorer

Only numbers stated in the episode are plotted. One run per model, no confidence intervals reported.

Two leakage channels

The versioned database closes only the first. Hover a box.

AEP case: 11 systems

Number of systems getting each step right. Heat = share of 11.

Web run: 49 of 86 trajectories

Astra touched post-cutoff sources in 57%, yet 68.1 (frozen) vs 68.3 (web). Hover cells.

Audits the episode suggests

Named in discussion as options; the paper is reported to run none of them.

References