feat: enhance model training and deployment scripts with time-weighted sampling

- Updated `train_model.py` and `train_mlx_model.py` to include a time weight decay parameter for improved sample weighting during training. - Modified dataset generation to incorporate sample weights based on time decay, enhancing model performance. - Adjusted deployment scripts to support new backend options and improved error handling for model file transfers. - Added new entries to the training log for better tracking of model performance metrics over time. - Included ONNX model export functionality in the MLX filter for compatibility with Linux servers.
2026-03-01 21:25:06 +09:00
parent 301457ce57
commit db144750a3
10 changed files with 324 additions and 97 deletions
--- a/models/training_log.json
+++ b/models/training_log.json
@@ -31,5 +31,19 @@
    "samples": 1696,
    "features": 21,
    "model_path": "models/lgbm_filter.pkl"
+  },
+  {
+    "date": "2026-03-01T21:03:56.314547",
+    "auc": 0.5406,
+    "samples": 1707,
+    "features": 21,
+    "model_path": "models/lgbm_filter.pkl"
+  },
+  {
+    "date": "2026-03-01T21:12:23.866860",
+    "auc": 0.502,
+    "samples": 3269,
+    "features": 21,
+    "model_path": "models/lgbm_filter.pkl"
  }
 ]
--- a/requirements.txt
+++ b/requirements.txt
@@ -12,3 +12,4 @@ lightgbm>=4.3.0
 scikit-learn>=1.4.0
 joblib>=1.3.0
 pyarrow>=15.0.0
+onnxruntime>=1.18.0
--- a/scripts/deploy_model.sh
+++ b/scripts/deploy_model.sh
@@ -1,65 +1,77 @@
 #!/usr/bin/env bash
 # 맥미니에서 학습한 모델을 LXC 컨테이너 볼륨 경로로 전송한다.
-# 사용법: bash scripts/deploy_model.sh [LXC_HOST] [LXC_MODELS_PATH]
+# 사용법: bash scripts/deploy_model.sh [lgbm|mlx]
 #
 # 예시:
-#   bash scripts/deploy_model.sh 10.1.10.28 /path/to/cointrader/models
-#   bash scripts/deploy_model.sh root@10.1.10.28 /root/cointrader/models
+#   bash scripts/deploy_model.sh        # LightGBM (기본값)
+#   bash scripts/deploy_model.sh mlx    # MLX 신경망

 set -euo pipefail

-LXC_HOST="${1:-root@10.1.10.24}"
-LXC_MODELS_PATH="${2:-/root/cointrader/models}"
-LOCAL_MODEL="models/lgbm_filter.pkl"
+BACKEND="${1:-lgbm}"
+LXC_HOST="root@10.1.10.24"
+LXC_MODELS_PATH="/root/cointrader/models"
 LOCAL_LOG="models/training_log.json"

-if [[ ! -f "$LOCAL_MODEL" ]]; then
-  echo "[오류] 모델 파일 없음: $LOCAL_MODEL"
-  echo "먼저 python scripts/train_model.py 를 실행하세요."
-  exit 1
+# ── 백엔드별 파일 목록 설정 ──────────────────────────────────────────────────
+# mlx: ONNX 파일만 전송 (Linux 서버는 onnxruntime으로 추론)
+# lgbm: pkl 파일 전송
+RELOAD_CMD="from src.ml_filter import MLFilter; f=MLFilter(); f.reload_model(); print('리로드 완료')"
+if [ "$BACKEND" = "mlx" ]; then
+  LOCAL_FILES=("models/mlx_filter.weights.onnx")
+else
+  LOCAL_FILES=("models/lgbm_filter.pkl")
 fi

-echo "=== 모델 전송 시작 ==="
-echo "  대상: ${LXC_HOST}:${LXC_MODELS_PATH}"
-echo "  파일: $LOCAL_MODEL"
-
-# 기존 모델을 prev로 백업 (원격)
-ssh "${LXC_HOST}" "
-  if [ -f '${LXC_MODELS_PATH}/lgbm_filter.pkl' ]; then
-    cp '${LXC_MODELS_PATH}/lgbm_filter.pkl' '${LXC_MODELS_PATH}/lgbm_filter_prev.pkl'
-    echo '  기존 모델 백업 완료'
+# ── 파일 존재 확인 ────────────────────────────────────────────────────────────
+for f in "${LOCAL_FILES[@]}"; do
+  if [[ ! -f "$f" ]]; then
+    echo "[오류] 모델 파일 없음: $f"
+    exit 1
  fi
+done
+
+echo "=== 모델 전송 시작 (백엔드: ${BACKEND}) ==="
+echo "  대상: ${LXC_HOST}:${LXC_MODELS_PATH}"
+
+# ── 원격 디렉터리 생성 + lgbm 기존 모델 백업 ─────────────────────────────────
+ssh "${LXC_HOST}" "
  mkdir -p '${LXC_MODELS_PATH}'
+  if [ '$BACKEND' = 'lgbm' ] && [ -f '${LXC_MODELS_PATH}/lgbm_filter.pkl' ]; then
+    cp '${LXC_MODELS_PATH}/lgbm_filter.pkl' '${LXC_MODELS_PATH}/lgbm_filter_prev.pkl'
+    echo '  기존 lgbm 모델 백업 완료'
+  fi
 "

-# 모델 파일 전송 (rsync 우선, 없으면 scp 폴백)
-if command -v rsync &>/dev/null && ssh "${LXC_HOST}" "command -v rsync" &>/dev/null; then
-  rsync -avz --progress \
-    "$LOCAL_MODEL" \
-    "${LXC_HOST}:${LXC_MODELS_PATH}/lgbm_filter.pkl"
-else
-  echo "  rsync 없음 → scp 사용"
-  scp "$LOCAL_MODEL" "${LXC_HOST}:${LXC_MODELS_PATH}/lgbm_filter.pkl"
-fi
-
-# 학습 로그도 함께 전송 (있을 경우)
-if [[ -f "$LOCAL_LOG" ]]; then
+# ── 파일 전송 헬퍼 (rsync 우선, scp 폴백) ────────────────────────────────────
+_send() {
+  local src="$1" dst="$2"
+  echo "  전송: $src → ${LXC_HOST}:$dst"
  if command -v rsync &>/dev/null && ssh "${LXC_HOST}" "command -v rsync" &>/dev/null; then
-    rsync -avz "$LOCAL_LOG" "${LXC_HOST}:${LXC_MODELS_PATH}/training_log.json"
+    rsync -avz --progress "$src" "${LXC_HOST}:$dst"
  else
-    scp "$LOCAL_LOG" "${LXC_HOST}:${LXC_MODELS_PATH}/training_log.json"
+    scp "$src" "${LXC_HOST}:$dst"
  fi
+}
+
+# ── 모델 파일 전송 ────────────────────────────────────────────────────────────
+for f in "${LOCAL_FILES[@]}"; do
+  _send "$f" "${LXC_MODELS_PATH}/$(basename "$f")"
+done
+
+# ── 학습 로그 전송 ────────────────────────────────────────────────────────────
+if [[ -f "$LOCAL_LOG" ]]; then
+  _send "$LOCAL_LOG" "${LXC_MODELS_PATH}/training_log.json"
  echo "  학습 로그 전송 완료"
 fi

 echo "=== 전송 완료 ==="
 echo ""

-# 봇 컨테이너가 실행 중이면 모델 핫리로드, 아니면 건너뜀
+# ── 핫리로드 ─────────────────────────────────────────────────────────────────
 echo "=== 핫리로드 시도 ==="
 if ssh "${LXC_HOST}" "docker inspect -f '{{.State.Running}}' cointrader 2>/dev/null | grep -q true"; then
-  ssh "${LXC_HOST}" "docker exec cointrader python -c \
-    \"from src.ml_filter import MLFilter; f=MLFilter(); f.reload_model(); print('리로드 완료')\""
+  ssh "${LXC_HOST}" "docker exec cointrader python -c \"${RELOAD_CMD}\""
  echo "=== 핫리로드 완료 ==="
 else
  echo "  cointrader 컨테이너가 실행 중이 아닙니다. 건너뜁니다."
--- a/scripts/train_and_deploy.sh
+++ b/scripts/train_and_deploy.sh
@@ -1,10 +1,10 @@
 #!/usr/bin/env bash
 # 맥미니에서 전체 학습 파이프라인을 실행하고 LXC로 배포한다.
-# 사용법: bash scripts/train_and_deploy.sh [LXC_HOST] [LXC_MODELS_PATH]
+# 사용법: bash scripts/train_and_deploy.sh [mlx|lgbm]
 #
 # 예시:
-#   bash scripts/train_and_deploy.sh
-#   bash scripts/train_and_deploy.sh root@10.1.10.24 /root/cointrader/models
+#   bash scripts/train_and_deploy.sh        # LightGBM (기본값)
+#   bash scripts/train_and_deploy.sh mlx    # MLX GPU 학습

 set -euo pipefail

@@ -19,37 +19,35 @@ else
    echo "경고: 가상환경을 찾을 수 없습니다 ($VENV_PATH). 시스템 Python을 사용합니다." >&2
 fi

-LXC_HOST="${1:-root@10.1.10.24}"
-LXC_MODELS_PATH="${2:-/root/cointrader/models}"
+BACKEND="${1:-lgbm}"

 cd "$PROJECT_ROOT"

-echo "=== [1/3] 데이터 수집 (XRP + BTC + ETH 3심볼) ==="
+echo "=== [1/3] 데이터 수집 (XRP + BTC + ETH 3심볼, 1년치) ==="
 python scripts/fetch_history.py \
    --symbols XRPUSDT BTCUSDT ETHUSDT \
    --interval 1m \
-    --days 90 \
+    --days 365 \
    --output data/xrpusdt_1m.parquet
 # 결과: data/combined_1m.parquet (타임스탬프 기준 병합)

 echo ""
 echo "=== [2/3] 모델 학습 (21개 피처: XRP 13 + BTC/ETH 상관관계 8) ==="
-# TRAIN_BACKEND=mlx 로 설정하면 Apple Silicon GPU(Metal)를 사용한다 (기본: lgbm)
-BACKEND="${TRAIN_BACKEND:-lgbm}"
+DECAY="${TIME_WEIGHT_DECAY:-2.0}"
 if [ "$BACKEND" = "mlx" ]; then
-    echo "  백엔드: MLX (Apple Silicon GPU)"
-    python scripts/train_mlx_model.py --data data/combined_1m.parquet
+    echo "  백엔드: MLX (Apple Silicon GPU), decay=${DECAY}"
+    python scripts/train_mlx_model.py --data data/combined_1m.parquet --decay "$DECAY"
 else
-    echo "  백엔드: LightGBM (CPU)"
-    python scripts/train_model.py --data data/combined_1m.parquet
+    echo "  백엔드: LightGBM (CPU), decay=${DECAY}"
+    python scripts/train_model.py --data data/combined_1m.parquet --decay "$DECAY"
 fi

 echo ""
 echo "=== [3/3] LXC 배포 ==="
-bash scripts/deploy_model.sh "$LXC_HOST" "$LXC_MODELS_PATH"
+bash scripts/deploy_model.sh "$BACKEND"

 echo ""
 echo "=== 전체 파이프라인 완료 ==="
 echo ""
 echo "봇 재시작이 필요하면:"
-echo "  ssh ${LXC_HOST} 'cd /root/cointrader && docker compose restart cointrader'"
+echo "  ssh root@10.1.10.24 'cd /root/cointrader && docker compose restart cointrader'"
--- a/scripts/train_mlx_model.py
+++ b/scripts/train_mlx_model.py
@@ -25,14 +25,14 @@ MLX_MODEL_PATH = Path("models/mlx_filter.weights")
 LOG_PATH = Path("models/training_log.json")


-def train_mlx(data_path: str) -> float:
+def train_mlx(data_path: str, time_weight_decay: float = 2.0) -> float:
    print(f"데이터 로드: {data_path}")
    df = pd.read_parquet(data_path)
    print(f"캔들 수: {len(df)}")

    print("\n데이터셋 생성 중...")
    t0 = time.perf_counter()
-    dataset = generate_dataset_vectorized(df)
+    dataset = generate_dataset_vectorized(df, time_weight_decay=time_weight_decay)
    t1 = time.perf_counter()
    print(f"데이터셋 생성 완료: {t1 - t0:.1f}초, {len(dataset)}개 샘플")

@@ -46,10 +46,30 @@ def train_mlx(data_path: str) -> float:

    X = dataset[FEATURE_COLS]
    y = dataset["label"]
+    w = dataset["sample_weight"].values

    split = int(len(X) * 0.8)
    X_train, X_val = X.iloc[:split], X.iloc[split:]
    y_train, y_val = y.iloc[:split], y.iloc[split:]
+    w_train = w[:split]
+
+    # --- 클래스 불균형 처리: 언더샘플링 (가중치 인덱스 보존) ---
+    pos_idx = np.where(y_train == 1)[0]
+    neg_idx = np.where(y_train == 0)[0]
+
+    if len(neg_idx) > len(pos_idx):
+        np.random.seed(42)
+        neg_idx = np.random.choice(neg_idx, size=len(pos_idx), replace=False)
+
+    balanced_idx = np.concatenate([pos_idx, neg_idx])
+    np.random.shuffle(balanced_idx)
+
+    X_train = X_train.iloc[balanced_idx]
+    y_train = y_train.iloc[balanced_idx]
+    w_train = w_train[balanced_idx]
+
+    print(f"\n언더샘플링 적용 후 학습 데이터: {len(X_train)}개 (양성={y_train.sum()}, 음성={(y_train==0).sum()})")
+    # --------------------------------------

    print("\nMLX 신경망 학습 시작 (GPU)...")
    t2 = time.perf_counter()
@@ -60,7 +80,7 @@ def train_mlx(data_path: str) -> float:
        epochs=100,
        batch_size=256,
    )
-    model.fit(X_train, y_train)
+    model.fit(X_train, y_train, sample_weight=w_train)
    t3 = time.perf_counter()
    print(f"학습 완료: {t3 - t2:.1f}초")

@@ -83,6 +103,7 @@ def train_mlx(data_path: str) -> float:
        "auc": round(auc, 4),
        "samples": len(dataset),
        "train_sec": round(t3 - t2, 1),
+        "time_weight_decay": time_weight_decay,
        "model_path": str(MLX_MODEL_PATH),
    })
    with open(LOG_PATH, "w") as f:
@@ -94,8 +115,12 @@ def train_mlx(data_path: str) -> float:
 def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("--data", default="data/xrpusdt_1m.parquet")
+    parser.add_argument(
+        "--decay", type=float, default=2.0,
+        help="시간 가중치 감쇠 강도 (0=균등, 2.0=최신이 ~7.4배 높음)",
+    )
    args = parser.parse_args()
-    train_mlx(args.data)
+    train_mlx(args.data, time_weight_decay=args.decay)


 if __name__ == "__main__":
--- a/scripts/train_model.py
+++ b/scripts/train_model.py
@@ -146,7 +146,7 @@ def generate_dataset(df: pd.DataFrame, n_jobs: int | None = None) -> pd.DataFram
    return pd.DataFrame(rows)


-def train(data_path: str):
+def train(data_path: str, time_weight_decay: float = 2.0):
    print(f"데이터 로드: {data_path}")
    df_raw = pd.read_parquet(data_path)
    print(f"캔들 수: {len(df_raw)}, 컬럼: {list(df_raw.columns)}")
@@ -169,7 +169,7 @@ def train(data_path: str):
    df = df_raw[base_cols].copy()

    print("데이터셋 생성 중...")
-    dataset = generate_dataset_vectorized(df, btc_df=btc_df, eth_df=eth_df)
+    dataset = generate_dataset_vectorized(df, btc_df=btc_df, eth_df=eth_df, time_weight_decay=time_weight_decay)

    if dataset.empty or "label" not in dataset.columns:
        raise ValueError(f"데이터셋 생성 실패: 샘플 0개. 위 오류 메시지를 확인하세요.")
@@ -183,10 +183,30 @@ def train(data_path: str):
    print(f"사용 피처: {len(actual_feature_cols)}개 {actual_feature_cols}")
    X = dataset[actual_feature_cols]
    y = dataset["label"]
+    w = dataset["sample_weight"].values

    split = int(len(X) * 0.8)
    X_train, X_val = X.iloc[:split], X.iloc[split:]
    y_train, y_val = y.iloc[:split], y.iloc[split:]
+    w_train = w[:split]
+
+    # --- 클래스 불균형 처리: 언더샘플링 (가중치 인덱스 보존) ---
+    pos_idx = np.where(y_train == 1)[0]
+    neg_idx = np.where(y_train == 0)[0]
+
+    if len(neg_idx) > len(pos_idx):
+        np.random.seed(42)
+        neg_idx = np.random.choice(neg_idx, size=len(pos_idx), replace=False)
+
+    balanced_idx = np.concatenate([pos_idx, neg_idx])
+    np.random.shuffle(balanced_idx)
+
+    X_train = X_train.iloc[balanced_idx]
+    y_train = y_train.iloc[balanced_idx]
+    w_train = w_train[balanced_idx]
+
+    print(f"\n언더샘플링 적용 후 학습 데이터: {len(X_train)}개 (양성={y_train.sum()}, 음성={(y_train==0).sum()})")
+    # --------------------------------------

    model = lgb.LGBMClassifier(
        n_estimators=300,
@@ -201,6 +221,7 @@ def train(data_path: str):
    )
    model.fit(
        X_train, y_train,
+        sample_weight=w_train,
        eval_set=[(X_val, y_val)],
        callbacks=[lgb.early_stopping(30, verbose=False), lgb.log_evaluation(50)],
    )
@@ -225,9 +246,11 @@ def train(data_path: str):
            log = json.load(f)
    log.append({
        "date": datetime.now().isoformat(),
+        "backend": "lgbm",
        "auc": round(auc, 4),
        "samples": len(dataset),
        "features": len(actual_feature_cols),
+        "time_weight_decay": time_weight_decay,
        "model_path": str(MODEL_PATH),
    })
    with open(LOG_PATH, "w") as f:
@@ -239,8 +262,12 @@ def train(data_path: str):
 def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("--data", default="data/xrpusdt_1m.parquet")
+    parser.add_argument(
+        "--decay", type=float, default=2.0,
+        help="시간 가중치 감쇠 강도 (0=균등, 2.0=최신이 ~7.4배 높음)",
+    )
    args = parser.parse_args()
-    train(args.data)
+    train(args.data, time_weight_decay=args.decay)


 if __name__ == "__main__":
--- a/src/dataset_builder.py
+++ b/src/dataset_builder.py
@@ -275,26 +275,24 @@ def _calc_labels_vectorized(
        fut_high = highs[idx + 1 : end]
        fut_low  = lows[idx + 1 : end]

-        label = None
+        label = 0  # 미도달(타임아웃) 시 실패로 간주
+
        for h, l in zip(fut_high, fut_low):
            if signal == "LONG":
-                if h >= tp:
-                    label = 1
-                    break
                if l <= sl:
                    label = 0
                    break
-            else:
-                if l <= tp:
+                if h >= tp:
                    label = 1
                    break
+            else:  # SHORT
                if h >= sl:
                    label = 0
                    break
+                if l <= tp:
+                    label = 1
+                    break

-        if label is None:
-            valid_mask.append(False)
-        else:
        labels.append(label)
        valid_mask.append(True)

@@ -305,11 +303,17 @@ def generate_dataset_vectorized(
    df: pd.DataFrame,
    btc_df: pd.DataFrame | None = None,
    eth_df: pd.DataFrame | None = None,
+    time_weight_decay: float = 0.0,
 ) -> pd.DataFrame:
    """
    전체 시계열을 1회 계산해 학습 데이터셋을 생성한다.
    기존 generate_dataset()의 drop-in 대체제.
    btc_df, eth_df가 제공되면 21개 피처로 확장한다.
+
+    time_weight_decay: 지수 감쇠 강도. 0이면 균등 가중치.
+        양수일수록 최신 샘플에 더 높은 가중치를 부여한다.
+        예) 2.0 → 최신 샘플이 가장 오래된 샘플보다 e^2 ≈ 7.4배 높은 가중치.
+        결과 DataFrame에 'sample_weight' 컬럼으로 포함된다.
    """
    print("  [1/3] 전체 시계열 지표 계산 (1회)...")
    d = _calc_indicators(df)
@@ -338,4 +342,17 @@ def generate_dataset_vectorized(
    feat_final = feat_all.iloc[final_idx][available_feature_cols].copy()
    feat_final["label"] = labels

-    return feat_final.reset_index(drop=True)
+    # 시간 가중치: 오래된 샘플 → 낮은 가중치, 최신 샘플 → 높은 가중치
+    n = len(feat_final)
+    if time_weight_decay > 0 and n > 1:
+        weights = np.exp(time_weight_decay * np.linspace(0.0, 1.0, n)).astype(np.float32)
+        weights /= weights.mean()  # 평균 1로 정규화해 학습률 스케일 유지
+        print(f"  시간 가중치 적용 (decay={time_weight_decay}): "
+              f"min={weights.min():.3f}, max={weights.max():.3f}")
+    else:
+        weights = np.ones(n, dtype=np.float32)
+
+    feat_final = feat_final.reset_index(drop=True)
+    feat_final["sample_weight"] = weights
+
+    return feat_final
--- a/src/label_builder.py
+++ b/src/label_builder.py
@@ -9,21 +9,17 @@ def build_labels(
    stop_loss: float,
    side: str,
 ) -> Optional[int]:
-    """
-    진입 이후 미래 캔들을 순서대로 확인해 TP/SL 도달 여부를 판단한다.
-    LONG: high >= TP → 1, low <= SL → 0
-    SHORT: low <= TP → 1, high >= SL → 0
-    둘 다 미도달 → None (학습 데이터에서 제외)
-    """
    for high, low in zip(future_highs, future_lows):
        if side == "LONG":
-            if high >= take_profit:
-                return 1
+            # 보수적 접근: 손절(SL)을 먼저 체크
            if low <= stop_loss:
                return 0
-        else:  # SHORT
-            if low <= take_profit:
+            if high >= take_profit:
                return 1
+        else:  # SHORT
+            # 보수적 접근: 손절(SL)을 먼저 체크
            if high >= stop_loss:
                return 0
+            if low <= take_profit:
+                return 1
    return None
--- a/src/ml_filter.py
+++ b/src/ml_filter.py
@@ -1,32 +1,63 @@
 from pathlib import Path
 import joblib
+import numpy as np
 import pandas as pd
 from loguru import logger

+from src.ml_features import FEATURE_COLS
+
+ONNX_MODEL_PATH = Path("models/mlx_filter.weights.onnx")
+LGBM_MODEL_PATH = Path("models/lgbm_filter.pkl")
+

 class MLFilter:
    """
-    LightGBM 모델을 로드하고 진입 여부를 판단한다.
-    모델 파일이 없으면 항상 진입을 허용한다 (폴백).
+    ML 필터. ONNX(MLX 신경망) 우선 로드, 없으면 LightGBM으로 폴백한다.
+    둘 다 없으면 항상 진입을 허용한다.
+
+    우선순위: ONNX > LightGBM > 폴백(항상 허용)
    """

-    def __init__(self, model_path: str = "models/lgbm_filter.pkl", threshold: float = 0.60):
-        self._model_path = Path(model_path)
+    def __init__(
+        self,
+        onnx_path: str = str(ONNX_MODEL_PATH),
+        lgbm_path: str = str(LGBM_MODEL_PATH),
+        threshold: float = 0.60,
+    ):
+        self._onnx_path = Path(onnx_path)
+        self._lgbm_path = Path(lgbm_path)
        self._threshold = threshold
-        self._model = None
+        self._onnx_session = None
+        self._lgbm_model = None
        self._try_load()

    def _try_load(self):
-        if self._model_path.exists():
+        # ONNX 우선 시도
+        if self._onnx_path.exists():
            try:
-                self._model = joblib.load(self._model_path)
-                logger.info(f"ML 필터 모델 로드 완료: {self._model_path}")
+                import onnxruntime as ort
+                self._onnx_session = ort.InferenceSession(
+                    str(self._onnx_path),
+                    providers=["CPUExecutionProvider"],
+                )
+                self._lgbm_model = None
+                logger.info(f"ML 필터 ONNX 모델 로드 완료: {self._onnx_path}")
+                return
            except Exception as e:
-                logger.warning(f"ML 필터 모델 로드 실패: {e}")
-                self._model = None
+                logger.warning(f"ONNX 모델 로드 실패: {e}")
+                self._onnx_session = None
+
+        # LightGBM 폴백
+        if self._lgbm_path.exists():
+            try:
+                self._lgbm_model = joblib.load(self._lgbm_path)
+                logger.info(f"ML 필터 LightGBM 모델 로드 완료: {self._lgbm_path}")
+            except Exception as e:
+                logger.warning(f"LightGBM 모델 로드 실패: {e}")
+                self._lgbm_model = None

    def is_model_loaded(self) -> bool:
-        return self._model is not None
+        return self._onnx_session is not None or self._lgbm_model is not None

    def should_enter(self, features: pd.Series) -> bool:
        """
@@ -36,8 +67,13 @@ class MLFilter:
        if not self.is_model_loaded():
            return True
        try:
+            if self._onnx_session is not None:
+                input_name = self._onnx_session.get_inputs()[0].name
+                X = features[FEATURE_COLS].values.astype(np.float32).reshape(1, -1)
+                proba = float(self._onnx_session.run(None, {input_name: X})[0][0])
+            else:
                X = features.to_frame().T
-            proba = self._model.predict_proba(X)[0][1]
+                proba = float(self._lgbm_model.predict_proba(X)[0][1])
            logger.debug(f"ML 필터 확률: {proba:.3f} (임계값: {self._threshold})")
            return bool(proba >= self._threshold)
        except Exception as e:
@@ -46,5 +82,7 @@ class MLFilter:

    def reload_model(self):
        """재학습 후 모델을 핫 리로드한다."""
+        self._onnx_session = None
+        self._lgbm_model = None
        self._try_load()
        logger.info("ML 필터 모델 리로드 완료")
--- a/src/mlx_filter.py
+++ b/src/mlx_filter.py
@@ -1,6 +1,7 @@
 """
 Apple MLX 기반 경량 신경망 필터.
 M4의 통합 GPU를 자동으로 활용한다.
+학습 후 ONNX로 export해 Linux 서버에서 onnxruntime으로 추론한다.
 """
 import numpy as np
 import pandas as pd
@@ -12,6 +13,83 @@ from pathlib import Path
 from src.ml_features import FEATURE_COLS


+def _export_onnx(
+    weights_npz: Path,
+    meta_npz: Path,
+    onnx_path: Path,
+) -> None:
+    """
+    MLX 가중치(.npz)를 읽어 ONNX 그래프로 변환한다.
+    네트워크 구조: fc1(ReLU) → dropout(추론 시 비활성) → fc2(ReLU) → fc3 → sigmoid
+    """
+    import onnx
+    from onnx import helper, TensorProto, numpy_helper
+
+    meta = np.load(meta_npz)
+    mean: np.ndarray = meta["mean"].astype(np.float32)
+    std: np.ndarray  = meta["std"].astype(np.float32)
+    input_dim  = int(meta["input_dim"])
+    hidden_dim = int(meta["hidden_dim"])
+
+    w = np.load(weights_npz)
+    # MLX save_weights 키 패턴: fc1.weight, fc1.bias, ...
+    fc1_w = w["fc1.weight"].astype(np.float32)   # (hidden, input)
+    fc1_b = w["fc1.bias"].astype(np.float32)
+    fc2_w = w["fc2.weight"].astype(np.float32)   # (hidden//2, hidden)
+    fc2_b = w["fc2.bias"].astype(np.float32)
+    fc3_w = w["fc3.weight"].astype(np.float32)   # (1, hidden//2)
+    fc3_b = w["fc3.bias"].astype(np.float32)
+
+    def _t(name: str, arr: np.ndarray) -> onnx.TensorProto:
+        return numpy_helper.from_array(arr, name=name)
+
+    initializers = [
+        _t("mean",  mean),
+        _t("std",   std),
+        _t("fc1_w", fc1_w),
+        _t("fc1_b", fc1_b),
+        _t("fc2_w", fc2_w),
+        _t("fc2_b", fc2_b),
+        _t("fc3_w", fc3_w),
+        _t("fc3_b", fc3_b),
+    ]
+
+    nodes = [
+        # 정규화: (x - mean) / std
+        helper.make_node("Sub",     ["X", "mean"],      ["x_sub"]),
+        helper.make_node("Div",     ["x_sub", "std"],   ["x_norm"]),
+        # fc1: x_norm @ fc1_w.T + fc1_b
+        helper.make_node("Gemm",    ["x_norm", "fc1_w", "fc1_b"], ["fc1_out"],
+                         transB=1),
+        helper.make_node("Relu",    ["fc1_out"],         ["relu1"]),
+        # fc2: relu1 @ fc2_w.T + fc2_b
+        helper.make_node("Gemm",    ["relu1",  "fc2_w", "fc2_b"], ["fc2_out"],
+                         transB=1),
+        helper.make_node("Relu",    ["fc2_out"],         ["relu2"]),
+        # fc3: relu2 @ fc3_w.T + fc3_b  → (N, 1)
+        helper.make_node("Gemm",    ["relu2",  "fc3_w", "fc3_b"], ["logits"],
+                         transB=1),
+        # sigmoid → (N, 1)
+        helper.make_node("Sigmoid", ["logits"],          ["proba_2d"]),
+        # squeeze: (N, 1) → (N,)
+        helper.make_node("Flatten", ["proba_2d"],        ["proba"], axis=0),
+    ]
+
+    graph = helper.make_graph(
+        nodes,
+        "mlx_filter",
+        inputs=[helper.make_tensor_value_info("X", TensorProto.FLOAT, [None, input_dim])],
+        outputs=[helper.make_tensor_value_info("proba", TensorProto.FLOAT, [None])],
+        initializer=initializers,
+    )
+    model_proto = helper.make_model(graph, opset_imports=[helper.make_opsetid("", 17)])
+    model_proto.ir_version = 8
+    onnx.checker.check_model(model_proto)
+    onnx_path.parent.mkdir(exist_ok=True)
+    onnx.save(model_proto, str(onnx_path))
+    print(f"  ONNX export 완료: {onnx_path}")
+
+
 class _Net(nn.Module):
    """3층 MLP 이진 분류기."""

@@ -53,7 +131,12 @@ class MLXFilter:
        self._std: np.ndarray | None = None
        self._trained = False

-    def fit(self, X: pd.DataFrame, y: pd.Series) -> "MLXFilter":
+    def fit(
+        self,
+        X: pd.DataFrame,
+        y: pd.Series,
+        sample_weight: np.ndarray | None = None,
+    ) -> "MLXFilter":
        X_np = X[FEATURE_COLS].values.astype(np.float32)
        y_np = y.values.astype(np.float32)

@@ -61,11 +144,20 @@ class MLXFilter:
        self._std = X_np.std(axis=0) + 1e-8
        X_np = (X_np - self._mean) / self._std

+        w_np = sample_weight.astype(np.float32) if sample_weight is not None else None
+
        optimizer = optim.Adam(learning_rate=self.lr)

-        def loss_fn(model: _Net, x: mx.array, y: mx.array) -> mx.array:
+        def loss_fn(
+            model: _Net, x: mx.array, y: mx.array, w: mx.array | None
+        ) -> mx.array:
            logits = model(x)
-            return nn.losses.binary_cross_entropy(logits, y, with_logits=True)
+            per_sample = nn.losses.binary_cross_entropy(
+                logits, y, with_logits=True, reduction="none"
+            )
+            if w is not None:
+                return (per_sample * w).sum() / w.sum()
+            return per_sample.mean()

        loss_and_grad = nn.value_and_grad(self._model, loss_fn)

@@ -78,7 +170,8 @@ class MLXFilter:
                batch_idx = idx[start : start + self.batch_size]
                x_batch = mx.array(X_np[batch_idx])
                y_batch = mx.array(y_np[batch_idx])
-                loss, grads = loss_and_grad(self._model, x_batch, y_batch)
+                w_batch = mx.array(w_np[batch_idx]) if w_np is not None else None
+                loss, grads = loss_and_grad(self._model, x_batch, y_batch, w_batch)
                optimizer.update(self._model, grads)
                mx.eval(self._model.parameters(), optimizer.state)
                epoch_loss += loss.item()
@@ -114,6 +207,12 @@ class MLXFilter:
            input_dim=np.array(self.input_dim),
            hidden_dim=np.array(self.hidden_dim),
        )
+        # ONNX export: Linux 서버에서 onnxruntime으로 추론하기 위해 변환
+        try:
+            onnx_path = path.with_suffix(".onnx")
+            _export_onnx(weights_path, meta_path, onnx_path)
+        except ImportError:
+            print("  [경고] onnx 패키지 없음 → ONNX export 생략 (pip install onnx)")

    @classmethod
    def load(cls, path: str | Path) -> "MLXFilter":