A public record.
Not a crystal ball.
An independent AI investigates market information, proposes a short-term portfolio plan, and keeps a short notebook. Application code validates every proposed order. The dashboard reads the resulting database records; a visit never triggers a trade.
What are we measuring?
Each experiment starts with its own capital. At the same start, we record a SPY price and calculate hypothetical shares: initial capital divided by that price. Those shares stay fixed. Their current value is the price-return benchmark; dividends are excluded. The difference in returns is expressed in percentage points.
The AI portfolio uses Alpaca account equity, cross-checked against cash plus holdings. Observations include timestamps. Missing or stale data is labeled; archived experiments keep their final marked-to-market results, including unsold holdings.
Paper execution and market data
This release only activates paper experiments. Paper fills are simulated and do not reproduce every feature of live execution, including market impact and dividends. Free IEX research data represents one exchange; simulated broker fills can use different market coverage. News and web research are evidence, not executable instructions.
Memory, not retraining
The notebook contains current hypotheses and tentative lessons. The agent can revise them using observed outcomes. It cannot modify application code, credentials, or fixed trading rules. New experiments inherit retained lessons with provenance, but start with fresh holdings and hypotheses. These are continuing experiments, not independent trials.
Decisions and outcomes
The agent explores opportunities over intraday to five trading sessions, aiming for positive returns and outperformance of SPY. A tick may propose up to three sequential orders; HOLD is valid and means no orders. An unfilled order stops the remaining proposals. A BUY proposal is not a purchase until the broker confirms a fill. Pending, partial, rejected, and failed outcomes remain visible. Short-term returns do not prove a strategy works or that memory improves investing.
Demo mode
Demo experiments contain synthetic UI fixtures. They make no brokerage or OpenAI calls and are always labeled. They must not be interpreted as actual model performance.
Costs and limits
API and hosting costs are separate from portfolio returns. Estimated API charges include conservative reservations for uncertain requests. The dashboard publishes selected effective settings. Capital and risk settings freeze at start; strategy policy changes are recorded with subsequent decisions. Hosted web-search context is provider-controlled: a request can exceed its token reservation before usage is reported. The application then stops the run before trading. Dollar limits are spending controls, not a guaranteed cap on the provider invoice.