main

ai-engine/scripts/train_v25_pro.py

@@ -101,6 +101,32 @@ FEATURES = [
     "home_top_scorer_form", "away_top_scorer_form",
     "home_avg_player_exp", "away_avg_player_exp",
     "home_goals_diversity", "away_goals_diversity",
+    # V27 H2H Expanded (4)
+    "h2h_home_goals_avg", "h2h_away_goals_avg",
+    "h2h_recent_trend", "h2h_venue_advantage",
+    # V27 Rolling Stats (13)
+    "home_rolling5_goals", "home_rolling5_conceded",
+    "home_rolling10_goals", "home_rolling10_conceded",
+    "home_rolling20_goals", "home_rolling20_conceded",
+    "away_rolling5_goals", "away_rolling5_conceded",
+    "away_rolling10_goals", "away_rolling10_conceded",
+    "home_rolling5_cs", "away_rolling5_cs",
+    # V27 Venue Stats (4)
+    "home_venue_goals", "home_venue_conceded",
+    "away_venue_goals", "away_venue_conceded",
+    # V27 Goal Trend (2)
+    "home_goal_trend", "away_goal_trend",
+    # V27 Calendar (5)
+    "home_days_rest", "away_days_rest",
+    "match_month", "is_season_start", "is_season_end",
+    # V27 Interaction (6)
+    "attack_vs_defense_home", "attack_vs_defense_away",
+    "xg_diff", "form_momentum_interaction",
+    "elo_form_consistency", "upset_x_elo_gap",
+    # V27 League Expanded (5)
+    "league_home_win_rate", "league_draw_rate",
+    "league_btts_rate", "league_ou25_rate",
+    "league_reliability_score",
 ]
 
 MARKET_CONFIGS = [
@@ -295,12 +321,18 @@ def train_market(df, target_col, market_name, num_class, n_trials):
 
     print(f"[INFO] Split: train={len(X_train)} val={len(X_val)} cal={len(X_cal)} test={len(X_test)}")
 
+    def _cb(study, trial):
+        if trial.number % 5 == 0 or trial.number == n_trials - 1:
+            best = study.best_value if study.best_trial else float('inf')
+            print(f"  [{trial.number+1:>3}/{n_trials}] loss={trial.value:.4f} | best={best:.4f}", flush=True)
+
     # ── Phase 1: Optuna XGBoost ──────────────────────────────────
     print(f"\n[OPTUNA] XGBoost tuning ({n_trials} trials)...")
     xgb_study = optuna.create_study(direction="minimize", sampler=TPESampler(seed=42))
     xgb_study.optimize(
         lambda trial: xgb_objective(trial, X_train, y_train, X_val, y_val, num_class),
         n_trials=n_trials,
+        callbacks=[_cb],
     )
     xgb_best = xgb_study.best_params
     print(f"[OK] XGB best logloss: {xgb_study.best_value:.4f}")
@@ -311,6 +343,7 @@ def train_market(df, target_col, market_name, num_class, n_trials):
     lgb_study.optimize(
         lambda trial: lgb_objective(trial, X_train, y_train, X_val, y_val, num_class),
         n_trials=n_trials,
+        callbacks=[_cb],
     )
     lgb_best = lgb_study.best_params
     print(f"[OK] LGB best logloss: {lgb_study.best_value:.4f}")