Investor's Last Exam: Can AI Agents Do Equity Investment Without Seeing the Future?

Investor's Last Exam. Can AI Agents Do Equity Investment Without Seeing the Future?. By Huangfei Jiang and nine others, from Prophetlab, RethinkLab, and four other institutions. Published on the twenty-eighth of October.

A financial agent has to form forecasts and investment judgments from the evidence available at the moment of decision, and those predictions are then judged against what actually happened next. Existing benchmarks assess parts of this: financial knowledge, research workflows, or forecasting. Evaluating the whole research process asks for two things together. A strictly frozen information set, so the agent cannot see the future. And professional criteria for what good research looks like.

Investor's Last Exam, or ILE, is an equity-research benchmark led by experienced buy-side investors and fund managers. Its evaluation spans eighty-six assignments and assesses three things jointly: the quality of the research, the accuracy of the forecasts, and simulated investment performance. Assessing them jointly is the point: it is the full research process that is being evaluated, from the evidence an analyst gathers to the judgment they finally make, rather than one skill at a time.

It has three components. The first is a versioned financial database spanning three point four three million document versions across nine source categories, supplemented by structured market data. Validated timestamps govern both retrieval and computation, which is how the information set stays frozen at decision time.

The second is a task taxonomy organized around the deliverables a professional analyst actually produces. The third is an expert-led scoring framework. It links each scoring criterion back to sources, calculations and assumptions through financial dependency graphs.

GPT-6 Astra and Claude Opus 5.5 lead overall, at sixty-eight point one and sixty-six point nine out of one hundred. Yet even Astra scores only fifty-two point five on Forecast and Model. The authors read this as evidence that even state-of-the-art models fail to translate overall research capability into predictive performance.

Put plainly, leading overall does not carry over into predicting well. The distance between Astra's overall score and its much lower score on Forecast and Model is where that shows, inside a single model. And because the information set is frozen at decision time, those forecasts are checked against outcomes the model could not have seen.

The shift ILE makes is in the unit being evaluated. Instead of testing isolated capabilities, it follows the chain from evidence to assumption to judgment. That is what lets a researcher locate where an analytical advantage stops carrying through into forecasts and investment decisions, and it gives a traceable, diagnostic base for training and evaluating financial agents.