Sprint 6-7 — ML Upgrade: Ensemble XGBoost+LightGBM+MLP + Optuna #1

admin · 2026-04-25T19:15:06+02:00

admin commented

2026-04-25 19:15:06 +02:00

Résumé

Implémentation de l'ensemble ML optimisé par Optuna pour les prédictions turf.

Résultats de performance

Modèle	Precision@3	AUC
Baseline XGBoost	0.5287	0.7254
Ensemble (voting)	0.5814	0.7840
Delta	+5.3%	—

✅ Seuil déploiement (+5%) atteint
✅ Latence : 35ms/course (< 200ms)
✅ 12/12 tests passants

Commits

feat(ml): add ensemble XGBoost+LightGBM+MLP with Optuna optimization — 100 trials par modèle
feat(ml): train ensemble model and generate benchmark report — rapport de benchmark

Lié à

HRT-32 Sprint 6-7 ML Upgrade

## Résumé Implémentation de l'ensemble ML optimisé par Optuna pour les prédictions turf. ## Résultats de performance | Modèle | Precision@3 | AUC | |--------|-------------|-----| | Baseline XGBoost | 0.5287 | 0.7254 | | **Ensemble (voting)** | **0.5814** | **0.7840** | | **Delta** | **+5.3%** | — | ✅ Seuil déploiement (+5%) atteint ✅ Latence : 35ms/course (< 200ms) ✅ 12/12 tests passants ## Commits - `feat(ml): add ensemble XGBoost+LightGBM+MLP with Optuna optimization` — 100 trials par modèle - `feat(ml): train ensemble model and generate benchmark report` — rapport de benchmark ## Lié à HRT-32 Sprint 6-7 ML Upgrade

admin added 2 commits 2026-04-25 19:15:07 +02:00

feat(ml): add ensemble XGBoost+LightGBM+MLP with Optuna optimization 0e7bcff6b0

- train_ensemble.py: full training pipeline with 100-trial Optuna studies
  for XGBoost and LightGBM, MLP (256-128-64), SHAP feature selection,
  weighted soft-voting ensemble, benchmark report generation
- predict_v2.py: production prediction module with model cache invalidation
- combined_api.py: add /api/v1/predictions, /api/v1/model/status,
  /api/v1/model/invalidate-cache endpoints using ensemble model
- tests/test_ml_ensemble.py: regression, latency and API tests

Baseline XGBoost Precision@3: 0.5287 (holdout 20% temporal)
Deploy threshold: +5% = 0.5551

Co-Authored-By: Paperclip <noreply@paperclip.ing>

feat(ml): train ensemble model and generate benchmark report 6b762068fd

Results:
  - XGBoost (Optuna 100 trials): AUC=0.7856, Precision@3=0.5783
  - LightGBM (Optuna 100 trials): AUC=0.7833, Precision@3=0.5736
  - MLP (3 layers 256-128-64): AUC=0.7743, Precision@3=0.5643
  - Ensemble (weighted voting): AUC=0.7840, Precision@3=0.5814

  Baseline XGBoost: Precision@3=0.5287
  Delta: +0.0527 (+5.3%) — DEPLOY threshold met (+5%)
  Latency: 35ms/race, 69ms/full-day (well under 200ms limit)

  SHAP: 31/43 features selected, top features: rang_cote,
  implied_prob, cote_direct, ratio_cote_field

  All 12 regression/latency tests passing.

Co-Authored-By: Paperclip <noreply@paperclip.ing>

admin merged commit f1ef2648b1 into master

2026-04-25 19:15:16 +02:00

admin referenced this issue from a commit

2026-04-25 19:15:17 +02:00

Merge pull request 'Sprint 6-7 — ML Upgrade: Ensemble XGBoost+LightGBM+MLP + Optuna' (#1) from feature/ml-upgrade-ensemble into master

admin referenced this issue from a commit

2026-05-24 09:22:17 +02:00

Fix #1: Ajout job run_ml_cache dans scheduler pour alimenter ml_predictions_cache

Sign in to join this conversation.

1 Participants

Notifications

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: admin/turf_saas#1