Add M1+M15 multi-timeframe SMC scalping training pipeline (GPU XGBoost),
then fix data-leakage and non-stationarity issues found in a skeptical audit.
Pipeline:
- src/triple_barrier.py: TP/SL/time labeling (ATR-scaled, asymmetric RR)
- src/multi_tf_dataset.py: M1 base + M15 HTF context, point-in-time join_asof
(only CLOSED M15 candles visible to each M1 bar - proven no leakage)
- src/economic_calendar.py: point-in-time forecast/actual/surprise provider
- src/smc_polars.py: add premium/discount + displacement SMC features
- scripts/train_multitf_scalper.py: GPU (device=cuda) training + walk-forward
- scripts/download_training_data.py: 1y data downloader
Leakage / robustness fixes (audit):
- CRITICAL: order block signal was written to the ORIGIN bar (future info);
now assigned at the CONFIRMATION bar -> matches live conditions
- replace non-stationary absolute features (ema_9/21, macd*) with scale-free
forms (ema*_dist_atr, ema_spread_atr, macd_*_bps) -> valid at any price level
- drop constant-zero calendar features from defaults (recurring provider has
no real values); re-add when a real calendar CSV is configured
- walk-forward + train/test now embargo the max_holding label horizon and drop
warmup rows (NaN->0 artifacts)
- news calendar features remain point-in-time (actual only at/after release)
Honest result: after fixes the spurious +2.35% edge collapses to ~random
(AUC 0.49). The prior edge was caused by the order-block look-ahead. Pipeline
is now leakage-free; a real edge still needs more M1 history / better features.
Also: test infra (pytest.ini asyncio, hmmlearn), TRAIN_BARS, cleanup of dead
modules. 14 tests pass.
Vanszs
a55148f232