Methodology
Written to read like a short research paper. If a claim on this site is not reproducible from this page, that's a bug — please tell us.
1. Data
- Public daily open / high / low / close / volume for ~50 large-cap US tickers.
- Source: Yahoo Finance public endpoints (yfinance library).
- No news, no sentiment, no alternative data, no non-public information.
- Corporate actions (splits, dividends) are adjusted into the price series.
2. Model
- 5-minute intraday bars are converted into daily movement-character features; stocks are grouped by movement character and one model is trained per group (model family v2.1).
- Training uses walk-forward validation — the model never sees data after the point it is asked to predict (no look-ahead).
- Every deployed model is versioned with a checksum; signals record the exact model version that produced them.
- Feature math is unit-tested against independent reference values; a look-ahead test in CI fails the build if any feature at day T uses data after day T.
3. Validation (what the headline numbers mean)
| Evaluation window | Out-of-sample holdout, finalized 2026-08-04 |
| Signals evaluated | 469,000 |
| Discrimination (AUROC) | 0.642 — the model ranks candidates better than random |
| Precision at 5% coverage | 58.4% — of the top 5% of signals, 58.4% reached the target |
| Net return per signal | +2.6% after estimated costs (commission + slippage) |
| Target | +3% move within 2 trading days |
These are backtested numbers. They describe what the model did on past data under stated assumptions. Past performance is not indicative of future results.
4. Costs assumed
- Commission: $0.005/share with a $1 minimum, plus $0.01/share estimated slippage on market orders at the close.
- No leverage, no shorting in the baseline evaluation.
5. Known limits
- Probabilities are estimates from historical data — the market changes and the estimate can be wrong.
- Models can drift: live performance is checked monthly against the honest backtest range; if it diverges materially, the model is taken offline publicly.
- Free data sources can be delayed or occasionally wrong; outcomes are recomputed from the same public data used at signal time.
- ~50 tickers is a small universe by design — large, liquid US names only.
6. Calibration
A monthly calibration report is published: "when the model said 60%, it happened X% of the time." If calibration breaks, you will see it before we do.