Researchers Build AI Framework to Automatically Generate Equity Research Reports from Financial Data

Financial institutions and analysts have long relied on detailed equity research reports to inform investment decisions. Now researchers are moving to automate this labor-intensive process using large language models, introducing the first systematic framework for generating these reports end-to-end.

The Challenge: Data and Metrics Gaps#

Creating a pipeline for equity research report (ERR) generation has proven difficult because of two fundamental obstacles. First, the field lacks publicly available datasets structured for training AI models on this specialized task. Second, there were no agreed-upon metrics to evaluate whether machine-generated reports are actually useful and accurate. This combination meant that despite LLMs' proven capability in related financial tasks—such as stock price forecasting and answering investor questions—no one had formalized the problem of full report generation.

FinRpt: A Benchmark Built from Multiple Data Sources#

To close this gap, researchers developed FinRpt, an open-source benchmark that combines seven different financial data types into a unified pipeline. The system automatically assembles a high-quality dataset suitable for both training and testing models. Rather than relying on manually written reports, the construction approach integrates diverse information streams to generate consistent training material.

Alongside the dataset, the team created a comprehensive evaluation system with 11 metrics to measure report quality and relevance. These metrics help distinguish between genuinely useful outputs and plausible-sounding but misleading analysis.

FinRpt-Gen: A Multi-Agent Architecture#

To tackle the generation task itself, the researchers introduced FinRpt-Gen, a multi-agent framework specifically designed for equity research report generation. The system leverages multiple LLM-based agents working in concert rather than relying on a single model. The team trained agents using two complementary approaches:

• Supervised Fine-Tuning: Learning directly from the FinRpt dataset to replicate report structure and content
• Reinforcement Learning: Optimizing performance based on the evaluation metrics to improve output quality over time