Most claims that "AI picks winning stocks" lack forward-only, auditable evidence. This project answers a narrower but crucial question: when constrained by preset rules and real-market data, can an LLM (ChatGPT) act as a transparent, auditable portfolio decision-maker and produce measurable results?
What Sets It Apart
- Forward-only live experiment with real money and immutable logs — every daily update, trade decision, and chat is preserved so results can't be retroactively altered, which makes causal claims testable.
- End-to-end research artifacts (trade logs, CSV accounting, charts, and a 40-page evaluation PDF) — so you get both raw data and analyzed metrics (Sharpe, Sortino, drawdown, CAPM) without reconstructing the run.
- Rules-driven LLM decision pipeline with automated enforcement (stop-loss, portfolio accounting) — the experiment separates model recommendations from enforced execution constraints, reducing covert human intervention.
- Reusable framework for replication (links to LIBB benchmark) — the repo doubles as a baseline for future LLM-investor behavior experiments rather than a one-off demo.
Who It's For and Tradeoffs
Great fit if you are a researcher or practitioner who wants reproducible evidence of LLM behavior in financial decision-making, or if you need a baseline framework to run comparable LLM-managed trading experiments. The project prioritizes transparency and auditability over production-grade robustness: it's an experimental research platform (Python, yfinance/stooq) not a regulated trading product, and micro-cap trading carries high market and liquidity risk. Expect that results reflect the experiment's constrained rules, data sources, and the specific ChatGPT configuration used — changing any of those elements can materially change outcomes.
Where It Fits
This sits between demonstration notebooks and full MLOps trading systems: more rigorous and auditable than a quick demo, but not a fully regulated, latency-optimized trading infrastructure. Use it to evaluate model decision patterns, failure modes, and replicability rather than to run high-frequency or institutional-scale strategies.
Methodology (brief)
The experiment runs daily cycles where the LLM receives research context and market data, proposes buy/sell decisions under hard constraints, and the repo records every step. Post-run evaluation compares portfolio performance versus benchmarks (S&P 500, Russell 2000) and computes risk metrics. Historical artifacts and chats are kept immutable to enable independent auditing and retrospective analysis.
