# Benchmark ML Ensemble — Turf Prédictions

**Date:** 2026-04-25  
**Dataset:** 10,899 partants  
**Holdout:** 2,180 lignes (2026-04-19 → 2026-04-24)

## Résultats

| Modèle | Precision@3 | AUC | Latence/prédiction |
|--------|-------------|-----|-------------------|
| XGBoost (baseline) | 0.5287 | 0.7254 | — |
| xgboost | 0.5783 | 0.7856 | 0.01 ms |
| lightgbm | 0.5736 | 0.7833 | 0.00 ms |
| mlp | 0.5643 | 0.7743 | 0.01 ms |
| **Ensemble** | **0.5814** | **0.7840** | **0.02 ms** |

## Décision de déploiement

- Delta Precision@3 : **+0.0527** (+5.3%)
- Seuil requis : **+5%**
- Résultat : **✅ DEPLOIEMENT RECOMMANDE**

## Optimisation Optuna

- Trials XGBoost : 100
- Trials LightGBM : 100
- Pruning : MedianPruner

### Meilleurs hyperparamètres XGBoost
```json
{
  "n_estimators": 141,
  "max_depth": 5,
  "learning_rate": 0.016298172447266404,
  "subsample": 0.7660470794373848,
  "colsample_bytree": 0.471124415020467,
  "min_child_weight": 14,
  "reg_alpha": 1.9364166463791586,
  "reg_lambda": 6.018030083488602,
  "gamma": 4.614943551368141
}
```

### Meilleurs hyperparamètres LightGBM
```json
{
  "n_estimators": 186,
  "max_depth": 4,
  "learning_rate": 0.012915117465216954,
  "num_leaves": 141,
  "subsample": 0.6193119116922561,
  "colsample_bytree": 0.539310022549326,
  "min_child_samples": 9,
  "reg_alpha": 0.6864583098112754,
  "reg_lambda": 0.0549259590914184
}
```

## Features

- Total features : 43
- Retenues par SHAP : 31

## Poids de l'ensemble

- xgboost : 0.2316
- lightgbm : 0.2342
- mlp : 0.2129