AI Agents in Equity Investment: Forecasting Without Future Insight

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

Authors: Huangfei Jiang, Anning Hu, Kai Sheng, Yuchen Cai, Rui Ge, Xiangyu Kong, Jianghao Lin, Shuzi Chen, Jiyi Weng, Junchi Yan

Executive Summary

Financial decision-makers, such as investors and fund managers, rely heavily on accurate forecasts and sound investment judgments based on real-time evidence. However, assessing the effectiveness of their research processes has proven challenging. Traditional benchmarks focus on specific aspects like financial knowledge or forecasting techniques but often miss the broader picture of how these elements work together in practice.

To address this gap, researchers have developed the Investor's Last Exam (ILE), a comprehensive benchmark designed to evaluate the entire equity-research process. This benchmark is led by seasoned buy-side investors and fund managers and consists of three key components:

1. Versioned Financial Database: The ILE includes a robust financial database with 3.43 million document versions across nine source categories. This database is supplemented by structured market data, all governed by validated timestamps for accurate retrieval and computation.

2. Task Taxonomy: The benchmark organizes tasks around the deliverables expected from professional analysts, ensuring that evaluations are relevant to the real-world responsibilities of financial agents.

3. Expert-Led Framework: An expert framework links scoring criteria to the sources, calculations, and assumptions used in analyses, employing financial dependency graphs to illustrate how various elements interact.

The evaluation process incorporates 86 assignments that assess three critical aspects: research quality, forecast accuracy, and simulated investment performance. Notably, the evaluations found that leading models, such as GPT-6 Astra and Claude Opus 5.5, scored 68.1 and 66.9 out of 100, respectively. However, even the highest-scoring model, Astra, received only a 52.5 in the Forecast & Model category. This indicates a significant gap; while these models show strong overall research capabilities, they struggle to convert this into effective predictive performance.

The implications of these findings are substantial. The ILE shifts the focus of evaluation from isolated analytical capabilities to the entire chain of evidence, assumptions, and judgments that drive investment decisions. This approach enables researchers and investors to pinpoint where analytical advantages fail to translate into actionable forecasts and investment outcomes.

For businesses, adopting a benchmark like the ILE could enhance the training and evaluation of financial agents, fostering a more systematic understanding of how research quality impacts predictive success. Organizations may find that using such a framework allows them to refine their investment strategies and improve the overall effectiveness of their research processes. The insights gained from the ILE could lead to better investment decisions, ultimately benefiting firms in a competitive financial landscape.

Academic Abstract

Financial agents must form forecasts and investment judgments from decision-time evidence, with predictions evaluated against subsequent outcomes. Existing benchmarks assess financial knowledge, research workflows, or forecasting, but evaluating the full research process requires a strictly frozen information set and professional evaluation criteria. We introduce Investor's Last Exam (ILE), an equity-research benchmark led by experienced buy-side investors and fund managers. Its three components are a versioned financial database spanning 3.43 million document versions and nine source categories, supplemented by structured market data, with validated timestamps governing retrieval and computation; a task taxonomy organized around professional analyst deliverables; and an expert-led framework linking scoring criteria to sources, calculations, and assumptions through financial dependency graphs. Evaluation spans 86 assignments and jointly assesses research quality, forecast accuracy, and simulated investment performance. GPT-6 Astra and Claude Opus 5.5 lead overall at 68.1 and 66.9 out of 100, yet even Astra scores only 52.5 on Forecast & Model. Our results show that even state-of-the-art models fail to translate overall research capability into predictive performance. ILE shifts the evaluation unit from isolated capabilities to the evidence--assumption--judgment chain, allowing researchers to locate where analytical advantages fail to translate into forecasts and investment judgments, and providing a traceable, diagnostic foundation for training and evaluating financial agents.