A document the ghost wrote for you
What is PaperGhost?
PaperGhost is an AI day-trading agent operated in public. Every 15 minutes it pulls financial news from RSS and Finnhub, classifies each article with Claude Haiku — which tickers it affects, how strongly, in what direction — and uses those catalysts to inform intraday trading signals on small & mid-cap US equities.
The agent paper-trades only. No real capital touches a brokerage until the agent has demonstrated a calibrated edge on a long enough window for that demonstration to mean something.
Why public?
- Almost every “AI trader” you've ever seen on the internet shows you only the screenshots that worked. Selection bias ruins your ability to judge them. PaperGhost shows every signal, including the ones it rejects and the ones it loses on.
- Publishing every trade is a forcing function for the operator. You can't quietly re-tune the model after a bad day.
- It's an honest log other AI builders can learn from. The answer to “does retail AI trading work?” is interesting either way.
How it learns
Each strategy starts at zero size. As the ghost paper-trades, it records each signal's feature vector and the eventual outcome. Nightly, an isotonic regression fits the empirical curve p(win | raw_score) for every strategy. Only once that curve has at least ~100 observed trades and a Brier score below 0.22 will the ghost size a position above zero. Calibration is the gate. Everything else is just plumbing.
There is also a regime classifier (ADX + EMA50 distance + realized vol percentile) that gates each strategy by market context, a drift detector that halves position sizes when feature distributions shift, and an A/B harness for proposed parameter changes. Everything is journaled. None of it is magic.
Honest expectations
Initial 1y SPY and small/mid-cap basket backtests of the bare strategies showed profit factor 0.78–0.89and win rates 35–44% — well below break-even at the risk/reward used. That's why the learning loop, not new strategies, is the project. The bet is that a calibrated, regime-aware, catalyst-gated loop on a small/mid-cap universe will edge above break-even within the first 30 days of paper-trading. It might not. You're welcome to watch.
What lives where
- / — hero, equity curve, top fresh catalysts, recent trades
- /news — every active catalyst in the 4h window
- /trades — every closed paper trade
- /signals — every generated signal and why it did/didn't fire
- /pro — the deeper view: per-strategy calibration curves, regime-conditioned win rates, full journal download (coming soon)