Merge pull request 'feat: OI nan 마스킹 / epsilon 통일 / 정밀도 우선 임계값' (#1) from feature/oi-nan-epsilon-precision-threshold into main

Reviewed-on: http://10.1.10.28:3000/gihyeon/cointrader/pulls/1

scripts/train_mlx_model.py

@@ -118,8 +118,26 @@ def train_mlx(data_path: str, time_weight_decay: float = 2.0) -> float:
 
     val_proba = model.predict_proba(X_val)
     auc = roc_auc_score(y_val, val_proba)
-    print(f"\n검증 AUC: {auc:.4f}")
+
-    print(classification_report(y_val, (val_proba >= 0.60).astype(int)))
+    # 최적 임계값 탐색: 최소 재현율(0.15) 조건부 정밀도 최대화
+    from sklearn.metrics import precision_recall_curve
+    precisions, recalls, thresholds = precision_recall_curve(y_val, val_proba)
+    precisions, recalls = precisions[:-1], recalls[:-1]
+
+    MIN_RECALL = 0.15
+    valid_idx = np.where(recalls >= MIN_RECALL)[0]
+    if len(valid_idx) > 0:
+        best_idx  = valid_idx[np.argmax(precisions[valid_idx])]
+        best_thr  = float(thresholds[best_idx])
+        best_prec = float(precisions[best_idx])
+        best_rec  = float(recalls[best_idx])
+    else:
+        best_thr, best_prec, best_rec = 0.50, 0.0, 0.0
+        print(f"  [경고] recall >= {MIN_RECALL} 조건 만족 임계값 없음 → 기본값 0.50 사용")
+
+    print(f"\n검증 AUC: {auc:.4f}  |  최적 임계값: {best_thr:.4f} "
+          f"(Precision={best_prec:.3f}, Recall={best_rec:.3f})")
+    print(classification_report(y_val, (val_proba >= best_thr).astype(int), zero_division=0))
 
     MLX_MODEL_PATH.parent.mkdir(exist_ok=True)
     model.save(MLX_MODEL_PATH)
@@ -133,6 +151,9 @@ def train_mlx(data_path: str, time_weight_decay: float = 2.0) -> float:
         "date": datetime.now().isoformat(),
         "backend": "mlx",
         "auc": round(auc, 4),
+        "best_threshold": round(best_thr, 4),
+        "best_precision": round(best_prec, 3),
+        "best_recall":    round(best_rec, 3),
         "samples": len(dataset),
         "train_sec": round(t3 - t2, 1),
         "time_weight_decay": time_weight_decay,

scripts/train_model.py

@@ -233,16 +233,26 @@ def train(data_path: str, time_weight_decay: float = 2.0):
 
     val_proba = model.predict_proba(X_val)[:, 1]
     auc = roc_auc_score(y_val, val_proba)
-    # 최적 임계값 탐색 (F1 기준)
+
-    thresholds = np.arange(0.40, 0.70, 0.05)
+    # 최적 임계값 탐색: 최소 재현율(0.15) 조건부 정밀도 최대화
-    best_thr, best_f1 = 0.50, 0.0
+    from sklearn.metrics import precision_recall_curve
-    for thr in thresholds:
+    precisions, recalls, thresholds = precision_recall_curve(y_val, val_proba)
-        pred = (val_proba >= thr).astype(int)
+    # precision_recall_curve의 마지막 원소는 (1.0, 0.0)이므로 제외
-        from sklearn.metrics import f1_score
+    precisions, recalls = precisions[:-1], recalls[:-1]
-        f1 = f1_score(y_val, pred, zero_division=0)
+
-        if f1 > best_f1:
+    MIN_RECALL = 0.15
-            best_f1, best_thr = f1, thr
+    valid_idx = np.where(recalls >= MIN_RECALL)[0]
-    print(f"\n검증 AUC: {auc:.4f}  |  최적 임계값: {best_thr:.2f} (F1={best_f1:.3f})")
+    if len(valid_idx) > 0:
+        best_idx  = valid_idx[np.argmax(precisions[valid_idx])]
+        best_thr  = float(thresholds[best_idx])
+        best_prec = float(precisions[best_idx])
+        best_rec  = float(recalls[best_idx])
+    else:
+        best_thr, best_prec, best_rec = 0.50, 0.0, 0.0
+        print(f"  [경고] recall >= {MIN_RECALL} 조건 만족 임계값 없음 → 기본값 0.50 사용")
+
+    print(f"\n검증 AUC: {auc:.4f}  |  최적 임계값: {best_thr:.4f} "
+          f"(Precision={best_prec:.3f}, Recall={best_rec:.3f})")
     print(classification_report(y_val, (val_proba >= best_thr).astype(int), zero_division=0))
 
     if MODEL_PATH.exists():
@@ -262,6 +272,9 @@ def train(data_path: str, time_weight_decay: float = 2.0):
         "date": datetime.now().isoformat(),
         "backend": "lgbm",
         "auc": round(auc, 4),
+        "best_threshold": round(best_thr, 4),
+        "best_precision": round(best_prec, 3),
+        "best_recall":    round(best_rec, 3),
         "samples": len(dataset),
         "features": len(actual_feature_cols),
         "time_weight_decay": time_weight_decay,

src/dataset_builder.py

@@ -116,12 +116,11 @@ def _calc_signals(d: pd.DataFrame) -> np.ndarray:
 
 
 def _rolling_zscore(arr: np.ndarray, window: int = 288) -> np.ndarray:
-    """rolling window z-score 정규화. 15분봉 기준 3일(288캔들) 윈도우.
+    """rolling window z-score 정규화. nan은 전파된다(nan-safe).
-    절대값 피처(ATR, 수익률 등)를 레짐 변화에 무관하게 만든다.
+    15분봉 기준 3일(288캔들) 윈도우. min_periods=1로 초반 데이터도 활용."""
-    min_periods=1로 초반 데이터도 활용하며, ddof=0(모표준편차)으로 계산한다."""
     s = pd.Series(arr.astype(np.float64))
     r = s.rolling(window=window, min_periods=1)
-    mean = r.mean()
+    mean = r.mean()   # pandas rolling은 nan을 자동으로 건너뜀
     std  = r.std(ddof=0)
     std  = std.where(std >= 1e-8, other=1e-8)
     z = (s - mean) / std
@@ -155,7 +154,7 @@ def _calc_features_vectorized(
     macd_sig = d["macd_signal"]
 
     bb_range = bb_upper - bb_lower
-    bb_pct = np.where(bb_range > 0, (close - bb_lower) / bb_range, 0.5)
+    bb_pct = (close - bb_lower) / (bb_range + 1e-8)
 
     ema_align = np.where(
         (ema9 > ema21) & (ema21 > ema50),  1,
@@ -164,8 +163,8 @@ def _calc_features_vectorized(
         )
     ).astype(np.float32)
 
-    atr_pct   = np.where(close > 0, atr / close, 0.0)
+    atr_pct   = atr / (close + 1e-8)
-    vol_ratio = np.where(vol_ma20 > 0, volume / vol_ma20, 1.0)
+    vol_ratio = volume / (vol_ma20 + 1e-8)
 
     ret_1 = close.pct_change(1).fillna(0).values
     ret_3 = close.pct_change(3).fillna(0).values
@@ -243,8 +242,8 @@ def _calc_features_vectorized(
         eth_r5 = _align(eth_ret_5, n).astype(np.float32)
 
         xrp_r1 = ret_1.astype(np.float32)
-        xrp_btc_rs_raw = np.where(btc_r1 != 0, xrp_r1 / btc_r1, 0.0).astype(np.float32)
+        xrp_btc_rs_raw = (xrp_r1 / (btc_r1 + 1e-8)).astype(np.float32)
-        xrp_eth_rs_raw = np.where(eth_r1 != 0, xrp_r1 / eth_r1, 0.0).astype(np.float32)
+        xrp_eth_rs_raw = (xrp_r1 / (eth_r1 + 1e-8)).astype(np.float32)
 
         extra = pd.DataFrame({
             "btc_ret_1":  _rolling_zscore(btc_r1),
@@ -258,10 +257,18 @@ def _calc_features_vectorized(
         }, index=d.index)
         result = pd.concat([result, extra], axis=1)
 
-    # OI 변화율 / 펀딩비 피처 (parquet에 컬럼이 있으면 z-score, 없으면 0)
+    # OI 변화율 / 펀딩비 피처
-    # OI는 최근 30일치만 제공되므로 이전 구간은 0으로 채워진 채로 들어옴
+    # 컬럼 없으면 전체 nan, 있으면 0.0 구간(데이터 미제공 구간)을 nan으로 마스킹
-    oi_raw = d["oi_change"].values if "oi_change" in d.columns else np.zeros(len(d))
+    # LightGBM은 nan을 자체 처리; MLX는 fit()에서 nanmean/nanstd + nan_to_num 처리
-    fr_raw = d["funding_rate"].values if "funding_rate" in d.columns else np.zeros(len(d))
+    if "oi_change" in d.columns:
+        oi_raw = np.where(d["oi_change"].values == 0.0, np.nan, d["oi_change"].values)
+    else:
+        oi_raw = np.full(len(d), np.nan)
+
+    if "funding_rate" in d.columns:
+        fr_raw = np.where(d["funding_rate"].values == 0.0, np.nan, d["funding_rate"].values)
+    else:
+        fr_raw = np.full(len(d), np.nan)
 
     result["oi_change"]    = _rolling_zscore(oi_raw.astype(np.float64))
     result["funding_rate"] = _rolling_zscore(fr_raw.astype(np.float64))
@@ -356,7 +363,12 @@ def generate_dataset_vectorized(
     feat_all   = _calc_features_vectorized(d, signal_arr, btc_df=btc_df, eth_df=eth_df)
 
     # 신호 발생 + NaN 없음 + 미래 데이터 충분한 인덱스만
-    available_cols_for_nan_check = [c for c in FEATURE_COLS if c in feat_all.columns]
+    # oi_change/funding_rate는 선택적 피처(컬럼 없으면 전체 nan)이므로 NaN 체크에서 제외
+    OPTIONAL_COLS = {"oi_change", "funding_rate"}
+    available_cols_for_nan_check = [
+        c for c in FEATURE_COLS
+        if c in feat_all.columns and c not in OPTIONAL_COLS
+    ]
     valid_rows = (
         (signal_arr != "HOLD") &
         (~feat_all[available_cols_for_nan_check].isna().any(axis=1).values) &

src/mlx_filter.py

@@ -140,9 +140,12 @@ class MLXFilter:
         X_np = X[FEATURE_COLS].values.astype(np.float32)
         y_np = y.values.astype(np.float32)
 
-        self._mean = X_np.mean(axis=0)
+        # nan-safe 정규화: nanmean/nanstd로 통계 계산 후 nan → 0.0 대치
-        self._std = X_np.std(axis=0) + 1e-8
+        # (z-score 후 0.0 = 평균값, 신경망에 줄 수 있는 가장 무난한 결측 대치값)
+        self._mean = np.nanmean(X_np, axis=0)
+        self._std  = np.nanstd(X_np, axis=0) + 1e-8
         X_np = (X_np - self._mean) / self._std
+        X_np = np.nan_to_num(X_np, nan=0.0)
 
         w_np = sample_weight.astype(np.float32) if sample_weight is not None else None
 
@@ -186,6 +189,7 @@ class MLXFilter:
         X_np = X[FEATURE_COLS].values.astype(np.float32)
         if self._trained and self._mean is not None:
             X_np = (X_np - self._mean) / self._std
+            X_np = np.nan_to_num(X_np, nan=0.0)
         x = mx.array(X_np)
         self._model.eval()
         logits = self._model(x)

tests/test_dataset_builder.py

@@ -91,3 +91,72 @@ def test_matches_original_generate_dataset(sample_df):
     assert 0.5 <= ratio <= 2.0, (
         f"샘플 수 차이가 너무 큼: 벡터화={len(vec)}, 기존={len(orig)}, 비율={ratio:.2f}"
     )
+
+
+def test_epsilon_no_division_by_zero():
+    """bb_range=0, close=0, vol_ma20=0 극단값에서 nan/inf가 발생하지 않아야 한다."""
+    import numpy as np
+    import pandas as pd
+    from src.dataset_builder import _calc_features_vectorized, _calc_signals, _calc_indicators
+
+    n = 100
+    # close를 모두 같은 값으로 → bb_range=0 유발
+    df = pd.DataFrame({
+        "open":   np.ones(n),
+        "high":   np.ones(n),
+        "low":    np.ones(n),
+        "close":  np.ones(n),
+        "volume": np.ones(n),
+    })
+    d = _calc_indicators(df)
+    sig = _calc_signals(d)
+    feat = _calc_features_vectorized(d, sig)
+
+    numeric_cols = feat.select_dtypes(include=[np.number]).columns
+    assert not feat[numeric_cols].isin([np.inf, -np.inf]).any().any(), \
+        "inf 값이 있으면 안 됨"
+
+
+def test_oi_nan_masking_no_column():
+    """oi_change 컬럼이 없으면 전체가 nan이어야 한다."""
+    import numpy as np
+    import pandas as pd
+    from src.dataset_builder import _calc_features_vectorized, _calc_signals, _calc_indicators
+
+    n = 100
+    np.random.seed(0)
+    df = pd.DataFrame({
+        "open":   np.random.uniform(1, 2, n),
+        "high":   np.random.uniform(2, 3, n),
+        "low":    np.random.uniform(0.5, 1, n),
+        "close":  np.random.uniform(1, 2, n),
+        "volume": np.random.uniform(1000, 5000, n),
+    })
+    d = _calc_indicators(df)
+    sig = _calc_signals(d)
+    feat = _calc_features_vectorized(d, sig)
+
+    assert feat["oi_change"].isna().all(), "oi_change 컬럼 없을 때 전부 nan이어야 함"
+
+
+def test_oi_nan_masking_with_zeros():
+    """oi_change 컬럼이 있어도 0.0 구간은 nan으로 마스킹되어야 한다."""
+    import numpy as np
+    import pandas as pd
+    from src.dataset_builder import _calc_features_vectorized, _calc_signals, _calc_indicators
+
+    n = 100
+    np.random.seed(0)
+    df = pd.DataFrame({
+        "open":      np.random.uniform(1, 2, n),
+        "high":      np.random.uniform(2, 3, n),
+        "low":       np.random.uniform(0.5, 1, n),
+        "close":     np.random.uniform(1, 2, n),
+        "volume":    np.random.uniform(1000, 5000, n),
+        "oi_change": np.concatenate([np.zeros(50), np.random.uniform(-0.1, 0.1, 50)]),
+    })
+    d = _calc_indicators(df)
+    sig = _calc_signals(d)
+    feat = _calc_features_vectorized(d, sig)
+
+    assert feat["oi_change"].iloc[50:].notna().any(), "실제 OI 값 구간에 유한값이 있어야 함"

tests/test_mlx_filter.py

@@ -65,6 +65,31 @@ def test_mlx_filter_fit_and_predict():
     assert np.all((proba >= 0.0) & (proba <= 1.0))
 
 
+def test_fit_with_nan_features():
+    """oi_change 피처에 nan이 포함된 경우 학습이 정상 완료되어야 한다."""
+    import numpy as np
+    import pandas as pd
+    from src.mlx_filter import MLXFilter
+    from src.ml_features import FEATURE_COLS
+
+    n = 300
+    np.random.seed(42)
+    X = pd.DataFrame(
+        np.random.randn(n, len(FEATURE_COLS)).astype(np.float32),
+        columns=FEATURE_COLS,
+    )
+    # oi_change 앞 절반을 nan으로
+    X["oi_change"] = np.where(np.arange(n) < n // 2, np.nan, X["oi_change"])
+    y = pd.Series((np.random.rand(n) > 0.5).astype(np.float32))
+
+    model = MLXFilter(input_dim=len(FEATURE_COLS), hidden_dim=32, epochs=3)
+    model.fit(X, y)  # nan 있어도 예외 없이 완료되어야 함
+
+    proba = model.predict_proba(X)
+    assert not np.any(np.isnan(proba)), "예측 확률에 nan이 없어야 함"
+    assert proba.min() >= 0.0 and proba.max() <= 1.0
+
+
 def test_mlx_filter_save_load(tmp_path):
     """저장 후 로드한 모델이 동일한 예측값을 반환해야 한다."""
     from src.mlx_filter import MLXFilter