Methodology
Leakage-safe features, a four-model ensemble, quantile bands, calibrated probabilities and walk-forward validation: the full forecasting methodology.
1. Point-in-time data
Every feature is computed strictly from bars available at the decision date. Nothing downstream of the cutoff enters the feature frame, so a forecast dated last March uses only what was knowable last March.
2. Feature frame
Ten features cover momentum (1, 5 and 20 day returns), realised volatility (20 and 60 day, annualised), trend position (price versus the 20 and 60 day averages), participation (volume surprise z-score), and stress (drawdown from the 60-day high, intraday range).
3. Ensemble
Four estimators vote on expected return: trend-following, momentum persistence, mean reversion, and volatility-discounted carry. Each carries a fixed weight; the standard deviation of their views becomes the model disagreement statistic.
4. Uncertainty band
The dispersion is a blend of short and medium horizon realised volatility, scaled by the square root of the horizon and widened when the ensemble disagrees. Quantiles are reported at the 5th, 25th, 50th, 75th and 95th percentiles in both return and price space.
5. Probability and confidence
Bullish probability is the mass of the distribution above zero, capped between 2% and 98%. Confidence blends the length of available history against the volatility regime and the ensemble disagreement — a wide, unstable, contested view scores low regardless of the headline number.
6. Attribution
Drivers are ranked by signed feature contribution so you can see whether a bullish call rests on trend position, momentum, or simply a quiet volatility regime.
7. Walk-forward validation
Forecasts are re-run across historical folds and scored out of sample: directional accuracy, 90% interval coverage, mean absolute error, bias and Brier score. Coverage materially below 90% means the bands are too tight and should be treated with caution.
8. Data provenance
This deployment runs on a reproducible synthetic market generator seeded per symbol, so the entire research workflow can be exercised deterministically. The engine accepts vendor bars unchanged — the feature, ensemble and validation layers do not change when a live feed is attached.
All output is quantitative research, not investment advice. Forecast bands describe model uncertainty only and exclude event risk that the feature frame cannot observe.