One-click full training pipeline · force-reruns bootstrap, auto-train, calibrators · reports each step
Runs every training path the brain has in sequence:
Historical Bootstrap (250d × 47 syms) →
Auto-Trainer (latest Stooq bars, daily) →
Continuous-Learner resolve (settles any pending journal entries) →
Calibrator + IsotonicCalibrator + RegimeCalibrator fits.
Expect ~30 seconds total. Network use: ~50 KB of CSV. Safe to re-run anytime.
📚 What each step does
1. Historical Bootstrap — pulls 250 daily bars × 47 symbols from Stooq, reconstructs 22-feature vectors per (sym, day), labels by next-day direction, feeds the global model + every calibrator. Expected output: ~11,000 training rows.
2. Auto-Trainer — pulls only the freshest Stooq bars and trains on rows we haven't seen yet (deduped by date). Picks up everything that happened since the bootstrap ran. Trains main model + per-horizon ensemble (short/mid/long).
3. Continuous-Learner resolve — walks the captured-but-unresolved journal, resolves any entries whose horizon has elapsed, trains on the freshly-labeled pairs.
4. Calibrators — refits Platt (Calibrator), Isotonic PAV (IsotonicCalibrator), and per-regime calibration (RegimeCalibrator) on the populated (predicted, actual) pools so the next prediction is honest immediately.