GenAI in Finance — AI-Augmented Financial Analytics
Portfolio Intelligence & SEC Document Intelligence System
5
Python systems
3
Streamlit versions
35 yrs
Fund history
9
LLMs benchmarked
AI Governance Framework
Every AI-assisted step includes a documented human validation checkpoint: dual-worksheet verification, source citation testing, distractor prompting, and ECE-based calibration analysis.
Problem
Financial analysts face a dual challenge: AI tools accelerate analytical throughput but introduce model risk — hallucination, overconfidence, and unverified assumptions — that can corrupt financial outputs if not explicitly managed. The challenge is not whether to use AI, but how to use it without compromising output integrity.
Approach
Built five Python analytics systems and three Streamlit dashboard versions, with generative AI integrated at the code scaffolding, modeling, and research synthesis stages. Every AI output was subject to a documented human validation checkpoint — dual-worksheet verification, source citation testing, and distractor prompting to probe model reliability.
Validation
WACC model outputs were cross-checked against Bloomberg-sourced cost of equity and debt benchmarks. SEC parsing outputs were verified against source HTML filings. Beta estimates were compared across yfinance, CAPM regression, and Fama-French 3-factor regression. ECE (expected calibration error) was computed for 9 LLMs to quantify overconfidence systematically.
Output
Five production Python systems: WACC modeling, CAPM/factor model analytics, SMA signal detection, SEC 10-K extraction pipeline, and a 35-year mutual fund analytics engine. Three client-grade Streamlit dashboards with rolling factor analytics and live yfinance data. A 25-slide AI governance research presentation benchmarking 9 LLMs.
Limitations
Live market data from yfinance introduces look-back bias if not handled carefully — all backtests use point-in-time data. LLM benchmarking results reflect model versions available at the time of testing; model behavior changes with version updates. The SEC parsing pipeline is optimized for 10-K HTML structure and may require updates for format changes.
Next Improvements
Extending the SEC pipeline to handle 10-Q and 8-K filings for continuous monitoring. Building a structured AI output audit log that records model version, prompt, and validation status for every AI-assisted analysis step. Adding a retrieval-augmented generation (RAG) layer to ground LLM outputs in verified financial data sources.