Framework
Cadrage fixe et protocole
Protocole affiché: one_shot_v4_gap_val3m_test1m. Les anciens protocoles restent consultables séparément.
Données
Sourcedata300 CSV/home/aalex/dev/quant_trading/data/data300
Cachedataset_processed_6h.pkl/home/aalex/dev/quant_trading/data/dataset_processed_6h.pkl
Période2019-09-08 12:00:00+00:00→ 2026-05-22 12:00:00+00:00
Rows / features9793 / 37bougies 6h
Input80 × 6h≈ 20 jours
Output40 × 6hrendement panier cumulé ≈ 10 jours
Features
Par crypto: ret = log(close).diff(), range = (high-low)/close, vol_chg = log1p(volume).diff(). Puis basket_ret = moyenne des *_ret.
ADA_ret ADA_range ADA_vol_chg BCH_ret BCH_range BCH_vol_chg BNB_ret BNB_range BNB_vol_chg BTC_ret BTC_range BTC_vol_chg DOT_ret DOT_range DOT_vol_chg ETH_ret ETH_range ETH_vol_chg LINK_ret LINK_range LINK_vol_chg SOL_ret SOL_range SOL_vol_chg SUI_ret SUI_range SUI_vol_chg XRP_ret XRP_range XRP_vol_chg basket_ret basket_vol_7d basket_vol_30d basket_trend_1d basket_trend_5d cross_asset_dispersion basket_drawdown_30d
Drift et sign-flip
Train uniquement. Val/Test restent réels pour mesurer le long/short sur marché non synthétique.
RègleOption BOption B: inversion uniquement des features directionnelles (*_ret et basket_ret) + target; *_range et *_vol_chg inchangés; validation/test inchangés
Features inversées11ADA_ret, BCH_ret, BNB_ret, BTC_ret, DOT_ret, ETH_ret, LINK_ret, SOL_ret, SUI_ret, XRP_ret, basket_ret
Non inverséesrange / vol_chg*_range, *_vol_chg
Walk-forward
12 folds: Train 12 mois, Val 3 mois précédents, Test 1 mois courants; tests sur 2025; holdout aveugle à partir de 2026-01-01
| Fold | Train | Validation | Test |
|---|---|---|---|
| 1 | 2023-10-01 00:00:00+00:00 → 2024-09-30 18:00:00+00:00 | 2024-10-01 00:00:00+00:00 → 2024-12-31 18:00:00+00:00 | 2025-01-01 00:00:00+00:00 → 2025-01-31 18:00:00+00:00 |
| 2 | 2023-11-01 00:00:00+00:00 → 2024-10-31 18:00:00+00:00 | 2024-11-01 00:00:00+00:00 → 2025-01-31 18:00:00+00:00 | 2025-02-01 00:00:00+00:00 → 2025-02-28 18:00:00+00:00 |
| 3 | 2023-12-01 00:00:00+00:00 → 2024-11-30 18:00:00+00:00 | 2024-12-01 00:00:00+00:00 → 2025-02-28 18:00:00+00:00 | 2025-03-01 00:00:00+00:00 → 2025-03-31 18:00:00+00:00 |
| 4 | 2024-01-01 00:00:00+00:00 → 2024-12-31 18:00:00+00:00 | 2025-01-01 00:00:00+00:00 → 2025-03-31 18:00:00+00:00 | 2025-04-01 00:00:00+00:00 → 2025-04-30 18:00:00+00:00 |
| 5 | 2024-02-01 00:00:00+00:00 → 2025-01-31 18:00:00+00:00 | 2025-02-01 00:00:00+00:00 → 2025-04-30 18:00:00+00:00 | 2025-05-01 00:00:00+00:00 → 2025-05-31 18:00:00+00:00 |
| 6 | 2024-03-01 00:00:00+00:00 → 2025-02-28 18:00:00+00:00 | 2025-03-01 00:00:00+00:00 → 2025-05-31 18:00:00+00:00 | 2025-06-01 00:00:00+00:00 → 2025-06-30 18:00:00+00:00 |
| 7 | 2024-04-01 00:00:00+00:00 → 2025-03-31 18:00:00+00:00 | 2025-04-01 00:00:00+00:00 → 2025-06-30 18:00:00+00:00 | 2025-07-01 00:00:00+00:00 → 2025-07-31 18:00:00+00:00 |
| 8 | 2024-05-01 00:00:00+00:00 → 2025-04-30 18:00:00+00:00 | 2025-05-01 00:00:00+00:00 → 2025-07-31 18:00:00+00:00 | 2025-08-01 00:00:00+00:00 → 2025-08-31 18:00:00+00:00 |
| 9 | 2024-06-01 00:00:00+00:00 → 2025-05-31 18:00:00+00:00 | 2025-06-01 00:00:00+00:00 → 2025-08-31 18:00:00+00:00 | 2025-09-01 00:00:00+00:00 → 2025-09-30 18:00:00+00:00 |
| 10 | 2024-07-01 00:00:00+00:00 → 2025-06-30 18:00:00+00:00 | 2025-07-01 00:00:00+00:00 → 2025-09-30 18:00:00+00:00 | 2025-10-01 00:00:00+00:00 → 2025-10-31 18:00:00+00:00 |
| 11 | 2024-08-01 00:00:00+00:00 → 2025-07-31 18:00:00+00:00 | 2025-08-01 00:00:00+00:00 → 2025-10-31 18:00:00+00:00 | 2025-11-01 00:00:00+00:00 → 2025-11-30 18:00:00+00:00 |
| 12 | 2024-09-01 00:00:00+00:00 → 2025-08-31 18:00:00+00:00 | 2025-09-01 00:00:00+00:00 → 2025-11-30 18:00:00+00:00 | 2025-12-01 00:00:00+00:00 → 2025-12-31 18:00:00+00:00 |
Métriques et sélection
SélectionVal DirAcc meanmoyenne sur folds, jamais Test
ConfirmationTest DirAcc / PNLreporting uniquement
PNL proxysign(pred) × ysans frais/slippage/funding
AlerteGap Test-Valdétecte instabilité / overfit de validation