feat: add --compare flag for OI derived features A/B comparison

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

scripts/train_model.py

@@ -422,6 +422,113 @@ def walk_forward_auc(
     print(f"  폴드별: {[round(a, 4) for a in aucs]}")
 
 
+def compare(data_path: str, time_weight_decay: float = 2.0, tuned_params_path: str | None = None):
+    """기존 피처 vs OI 파생 피처 추가 버전 A/B 비교."""
+    import warnings
+
+    print("=" * 70)
+    print("  OI 파생 피처 A/B 비교 (30일 데이터 기반, 방향성 참고용)")
+    print("=" * 70)
+
+    df_raw = pd.read_parquet(data_path)
+    base_cols = ["open", "high", "low", "close", "volume"]
+    btc_df = eth_df = None
+    if "close_btc" in df_raw.columns:
+        btc_df = df_raw[[c + "_btc" for c in base_cols]].copy()
+        btc_df.columns = base_cols
+    if "close_eth" in df_raw.columns:
+        eth_df = df_raw[[c + "_eth" for c in base_cols]].copy()
+        eth_df.columns = base_cols
+    df = df_raw[base_cols].copy()
+    if "oi_change" in df_raw.columns:
+        df["oi_change"] = df_raw["oi_change"]
+    if "funding_rate" in df_raw.columns:
+        df["funding_rate"] = df_raw["funding_rate"]
+
+    dataset = generate_dataset_vectorized(
+        df, btc_df=btc_df, eth_df=eth_df,
+        time_weight_decay=time_weight_decay,
+        negative_ratio=5,
+    )
+
+    if dataset.empty:
+        raise ValueError("데이터셋 생성 실패")
+
+    lgbm_params, weight_scale = _load_lgbm_params(tuned_params_path)
+
+    # Baseline: OI 파생 피처 제외
+    BASELINE_EXCLUDE = {"oi_change_ma5", "oi_price_spread"}
+    baseline_cols = [c for c in FEATURE_COLS if c in dataset.columns and c not in BASELINE_EXCLUDE]
+    new_cols = [c for c in FEATURE_COLS if c in dataset.columns]
+
+    results = {}
+    for label, cols in [("Baseline", baseline_cols), ("New", new_cols)]:
+        X = dataset[cols]
+        y = dataset["label"]
+        w = dataset["sample_weight"].values
+        source = dataset["source"].values if "source" in dataset.columns else np.full(len(X), "signal")
+
+        split = int(len(X) * 0.8)
+        X_tr, X_val = X.iloc[:split], X.iloc[split:]
+        y_tr, y_val = y.iloc[:split], y.iloc[split:]
+        w_tr = (w[:split] * weight_scale).astype(np.float32)
+        source_tr = source[:split]
+
+        balanced_idx = stratified_undersample(y_tr.values, source_tr, seed=42)
+        X_tr_b = X_tr.iloc[balanced_idx]
+        y_tr_b = y_tr.iloc[balanced_idx]
+        w_tr_b = w_tr[balanced_idx]
+
+        model = lgb.LGBMClassifier(**lgbm_params, random_state=42, verbose=-1)
+        with warnings.catch_warnings():
+            warnings.simplefilter("ignore")
+            model.fit(X_tr_b, y_tr_b, sample_weight=w_tr_b)
+
+        proba = model.predict_proba(X_val)[:, 1]
+        auc = roc_auc_score(y_val, proba) if len(np.unique(y_val)) > 1 else 0.5
+
+        precs, recs, thrs = precision_recall_curve(y_val, proba)
+        precs, recs = precs[:-1], recs[:-1]
+        valid_idx = np.where(recs >= 0.15)[0]
+        if len(valid_idx) > 0:
+            best_i = valid_idx[np.argmax(precs[valid_idx])]
+            thr, prec, rec = float(thrs[best_i]), float(precs[best_i]), float(recs[best_i])
+        else:
+            thr, prec, rec = 0.50, 0.0, 0.0
+
+        # Feature importance
+        imp = dict(zip(cols, model.feature_importances_))
+        top10 = sorted(imp.items(), key=lambda x: x[1], reverse=True)[:10]
+
+        results[label] = {
+            "auc": auc, "precision": prec, "recall": rec,
+            "threshold": thr, "n_val": len(y_val),
+            "n_val_pos": int(y_val.sum()), "top10": top10,
+        }
+
+    # 비교 테이블 출력
+    n_base = len(baseline_cols)
+    n_new = len(new_cols)
+    print(f"\n{'지표':<20} {f'Baseline({n_base})':>15} {f'New({n_new})':>15} {'Delta':>10}")
+    print("-" * 62)
+    for metric in ["auc", "precision", "recall", "threshold"]:
+        b = results["Baseline"][metric]
+        n = results["New"][metric]
+        d = n - b
+        sign = "+" if d > 0 else ""
+        print(f"{metric:<20} {b:>15.4f} {n:>15.4f} {sign}{d:>9.4f}")
+
+    n_val = results["Baseline"]["n_val"]
+    n_pos = results["Baseline"]["n_val_pos"]
+    print(f"\n검증셋: n={n_val} (양성={n_pos}, 음성={n_val - n_pos})")
+    print("⚠ 30일 데이터 기반 — 방향성 참고용\n")
+
+    print("Feature Importance Top 10 (New):")
+    for feat_name, imp_val in results["New"]["top10"]:
+        marker = " ← NEW" if feat_name in BASELINE_EXCLUDE else ""
+        print(f"  {feat_name:<25} {imp_val:>6}{marker}")
+
+
 def main():
     parser = argparse.ArgumentParser()
     parser.add_argument("--data", default="data/combined_15m.parquet")
@@ -435,9 +542,13 @@ def main():
         "--tuned-params", type=str, default=None,
         help="Optuna 튜닝 결과 JSON 경로 (지정 시 기본 파라미터를 덮어씀)",
     )
+    parser.add_argument("--compare", action="store_true",
+                        help="OI 파생 피처 추가 전후 A/B 성능 비교")
     args = parser.parse_args()
 
-    if args.wf:
+    if args.compare:
+        compare(args.data, time_weight_decay=args.decay, tuned_params_path=args.tuned_params)
+    elif args.wf:
         walk_forward_auc(
             args.data,
             time_weight_decay=args.decay,