IPO Finance Agent: Benchmark of LLM Financial Analysts Beyond Finance Agent v2, with Automated Rubric Generation, on the SpaceX (SPCX) IPO

Battle Count: 50

The paper focuses on IPO due diligence and financial document analysis rather than quantitative trading strategies. While understanding IPO filings and valuation could inform pre-IPO investment decisions or event-driven trading around IPOs, the paper does not address trading signals, portfolio construction, market microstructure, or algorithmic execution. Its primary relevance is to fundamental analysis and research workflows rather than quantitative trading systems.

The system involves multiple complex components: (1) an agentic tool-use harness with web search and EDGAR access; (2) contextual retrieval requiring LLM-generated document-aware descriptions for each chunk before embedding; (3) a multi-stage automated rubric pipeline with fact extraction, agreement-based consolidation, rubric induction, iterative quality evaluation with three routing paths (stop/repair/enrich), deduplication, and numeric-payload-preservation checks; (4) ensemble of 5+ model answers for fact extraction; (5) human expert review step. The pipeline requires careful prompt engineering, model orchestration, and quality control at multiple stages.

Code and 70 public questions released on GitHub. However, 930 of 1000 questions are held private to prevent benchmark contamination. The contextual retrieval pipeline and rubric generation pipeline are described but full reproduction requires access to the private dataset. Sponsored API access for GLM models limits cost reproducibility. No ablation study provided to isolate component contributions.