Building an autonomous financial analyst with o1 and o3-mini
Endex is developing an autonomous AI Analyst that uses OpenAI's reasoning models to retrieve, synthesize, and reason through complex financial data, moving beyond simple search results to provide structured thinking and deep analysis for investment professionals.
Precision and Reasoning in Financial Workflows
Endex replaces traditional retrieval-augmented generation (RAG) with agents that use OpenAI's reasoning models to reflect on data, identify inconsistencies, and contextualize metrics. This approach ensures the precision required for high-stakes financial decision-making, where missing a single adjustment in EBITDA reconciliation or a "change in control provision" clause can significantly alter an investment outlook.
Autonomous agents can process financial reports, market data, and firm-specific knowledge to execute the following tasks:
- Precedent transaction overviews: Summarizing past deals to benchmark current opportunities.
- Earnings performance summaries: Analyzing quarterly and annual results.
- Investment committee (IC) memo preparation: Drafting the core documentation for investment decisions.
- Data room due diligence: Reviewing confidential documents for risk and opportunity.
These agents deliver outputs in formats such as emails, documents, Excel models, and slide decks, allowing analysts to trace conclusions back to their original sources.
Technical Performance and Automation Gains
Integrating OpenAI's o-series models has enabled Endex to achieve higher accuracy and efficiency in several key areas:
Multistep Financial Reasoning
OpenAI o1 has simplified the process of complex financial reasoning. Previously, Endex relied on chained completions and multiple verification steps; o1's inherent reasoning capabilities allow for more streamlined workflows without sacrificing accuracy.
Latency and Efficiency
By utilizing OpenAI o3-mini, Endex has reduced latency per turn to one-third of previous levels. This efficiency enables the detailed multi-step workflows required for analyzing confidential information packages (CIPs) and automating financial model reconciliation.
AI-Powered Cross-Checking
Agents can now identify discrepancies in financial data, such as flagging restatements in footnotes and surfacing inconsistencies with targeted citations, reducing the manual verification burden on human analysts.
Multimodal Analysis
Using the Finance Agent Retrieval (FAR) benchmark to measure context usage on tabular and chart data, Endex utilizes o1's vision capabilities to process investor presentations, internal decks, 8-Ks, and Excel models.
Evaluation and Model Optimization
Endex and OpenAI collaborated to develop a rigorous evaluation and testing framework to track metrics such as response latency, first-token generation time, and reasoning depth.
Key findings from this framework include:
- Expert Preference: In blind user testing, financial experts preferred responses generated by OpenAI's o1 model 70% of the time over non-reasoning models.
- Reinforcement Learning: Endex used reinforcement fine-tuning on GPT-4o mini and the o1 series to improve entity extraction and query intent mapping, specifically for research-intensive tasks like precedent transaction analysis.
"Our collaboration with OpenAI has unlocked tools to tailor model behavior to the reasoning and output style professionals expect," says Pratham Soni, co-founder at Endex.