Investor's Last Exam scores 11 AI models on 86 equity research tasks

An AI agent can earn full marks for its reasoning on a stock and still miss every forecast.

Shanghai Jiao Tong University researchers released Investor's Last Exam on 27 October 2026. The benchmark tests 11 AI models on 86 equity research tasks, scored end to end against expert rubrics.

The benchmark freezes the clock. Every search, document fetch and database query sees only data available before the task's cutoff date. The corpus holds 3,430,336 document versions and 1.52 terabytes of data, dated through 15 October 2026. It covers financial reports, earnings calls, analyst estimates, holdings data and news. Prophetlab led the work with Foresight Fund Management, IvyRock Asset Management and researchers from Shanghai Jiao Tong University.

The 86 assignments cover six research categories. These are discovery, evidence, forecasting, valuation, monitoring, and integrated reports. Eight more tasks cover market trading. Each category tracks a buy-side analyst's work. Expert-written rubrics score both the supporting analysis and the forecast outcome.

GPT-6 Astra scored 68.1 out of 100 overall, the highest of the 11 systems. Claude Opus 5.5 came second at 66.9, with DeepSeek V4.1 Flash third at 63.0. The ranking changed on a different measure, realised forecast outcomes. That score shows how close each forecast was to the real result. DeepSeek V4.1 Flash led that measure at 40.3. Hunyuan 4 Preview was second at 37.9, GPT-6 Astra third at 37.7.

Prophetlab also ran a trading simulation on 20 opportunities across 11 companies, from March to August 2026. Hunyuan 4 Preview, sixth overall, led the simulation with a cumulative return of 8.50 percent. Qwen 3.8 Max, fifth overall, finished tenth at 1.67 percent. GPT-6 Astra placed third at 7.18 percent. The always-long baseline returned 1.18 percent. The paper applies 10 basis points per trade leg and three percent annualised short-borrowing cost.

The paper warns that 20 opportunities over six months is a small sample. That size is too small to show persistent returns. The time-frozen tools cover an agent's data access after the cutoff. The paper flags pretraining knowledge as a separate limit. A Tesla task from June 2026 showed the risk. Three models each earned 50 of 50 supporting-analysis points but 0 of 50 on the forecast. They missed deliveries, revenue and margin by more than the zero-credit threshold.

The benchmark separates three things that often get read as one. A fluent research report, an accurate forecast and a profitable trade measure different work. An AI team building a finance product can read each score on its own. The next step is testing the model on the work its users will act on.

Source

Investor's Last Exam, Prophetlab, GitHub · Paper PDF · Project page