수정본

backend/app/api/endpoints/ml_ensemble.py

@@ -0,0 +1,170 @@
+"""
+ML 앙상블 API 엔드포인트
+
+- POST /train: 모델 학습
+- GET /predict/{stock_code}: 개별 종목 예측
+- GET /status: 모델 상태 조회
+- GET /backtest: 백테스트 결과
+"""
+
+from fastapi import APIRouter, HTTPException, Query, BackgroundTasks
+import logging
+import asyncio
+
+from app.services.ml_ensemble import get_ensemble
+from app.services.korea_investment_api import korea_investment_api
+from app.core.config import settings
+
+router = APIRouter()
+logger = logging.getLogger(__name__)
+
+
+@router.get("/status")
+async def get_model_status():
+    """앙상블 모델 상태 조회"""
+    ensemble = get_ensemble()
+    status = ensemble.get_status()
+    status["config"] = {
+        "enabled": settings.ML_ENSEMBLE_ENABLED,
+        "rule_weight": settings.ML_ENSEMBLE_RULE_WEIGHT,
+        "ml_weight": settings.ML_ENSEMBLE_ML_WEIGHT,
+        "buy_threshold": settings.ML_ENSEMBLE_BUY_THRESHOLD,
+        "confidence_min": settings.ML_ENSEMBLE_CONFIDENCE_MIN,
+    }
+    return status
+
+
+@router.get("/predict/{stock_code}")
+async def predict_stock(stock_code: str):
+    """개별 종목 ML 앙상블 예측"""
+    ensemble = get_ensemble()
+
+    if not ensemble.is_trained:
+        raise HTTPException(
+            status_code=503,
+            detail="모델이 학습되지 않았습니다. POST /ml-ensemble/train 으로 학습을 먼저 실행하세요."
+        )
+
+    try:
+        chart_resp = await korea_investment_api.get_daily_chart(stock_code, "D")
+        chart_data = chart_resp.get("output", [])
+
+        if not chart_data or len(chart_data) < 61:
+            raise HTTPException(status_code=400, detail=f"차트 데이터 부족: {len(chart_data) if chart_data else 0}일 (최소 61일 필요)")
+
+        result = ensemble.predict(chart_data)
+        result["stock_code"] = stock_code
+
+        return result
+
+    except HTTPException:
+        raise
+    except Exception as e:
+        logger.error(f"예측 실패 ({stock_code}): {e}")
+        raise HTTPException(status_code=500, detail=str(e))
+
+
+async def _train_task(limit: int):
+    """백그라운드 학습 작업"""
+    import numpy as np
+    from app.services.feature_engineering import (
+        chart_data_to_dataframe, extract_features, prepare_training_data
+    )
+    from app.services.korea_investment_api import KoreaInvestmentAPI
+
+    kis_api = KoreaInvestmentAPI()
+
+    # 종목 조회
+    stocks = []
+    try:
+        kospi = await kis_api.get_volume_rank("J")
+        if kospi and kospi.get("output"):
+            for item in kospi["output"]:
+                name = item.get("hts_kor_isnm", "")
+                skip = ["ETF", "레버리지", "인버스", "KODEX", "선물", "TIGER", "ARIRANG"]
+                if not any(k in name for k in skip):
+                    stocks.append((item["mksc_shrn_iscd"], name))
+    except Exception as e:
+        logger.error(f"종목 조회 실패: {e}")
+
+    await asyncio.sleep(3)
+
+    try:
+        kosdaq = await kis_api.get_volume_rank("Q")
+        if kosdaq and kosdaq.get("output"):
+            for item in kosdaq["output"]:
+                name = item.get("hts_kor_isnm", "")
+                skip = ["ETF", "레버리지", "인버스", "KODEX", "선물", "TIGER", "ARIRANG"]
+                if not any(k in name for k in skip):
+                    stocks.append((item["mksc_shrn_iscd"], name))
+    except Exception as e:
+        logger.error(f"종목 조회 실패: {e}")
+
+    stocks = stocks[:limit]
+    logger.info(f"학습 대상: {len(stocks)}개 종목")
+
+    # 데이터 수집
+    all_X_tab, all_X_seq, all_y = [], [], []
+    for i, (code, name) in enumerate(stocks):
+        try:
+            chart_resp = await kis_api.get_daily_chart(code, "D")
+            chart_data = chart_resp.get("output", [])
+            if not chart_data or len(chart_data) < 80:
+                continue
+
+            df = chart_data_to_dataframe(chart_data)
+            feat = extract_features(df)
+            X_tab, X_seq, y, _ = prepare_training_data(df, feat)
+            if len(X_tab) > 0:
+                all_X_tab.append(X_tab)
+                all_X_seq.append(X_seq)
+                all_y.append(y)
+                logger.info(f"[{i+1}/{len(stocks)}] {code}: {len(X_tab)}개 샘플")
+        except Exception as e:
+            logger.warning(f"[{i+1}] {code} 실패: {e}")
+        await asyncio.sleep(3)
+
+    if not all_X_tab:
+        logger.error("학습 데이터 없음")
+        return
+
+    X_tabular = np.vstack(all_X_tab)
+    X_seq = np.vstack(all_X_seq)
+    y = np.concatenate(all_y)
+
+    logger.info(f"총 {len(X_tabular)}개 샘플로 학습 시작")
+
+    ensemble = get_ensemble()
+    metrics = ensemble.train(X_tabular, X_seq, y)
+    logger.info(f"학습 완료: {metrics}")
+
+
+@router.post("/train")
+async def train_model(
+    background_tasks: BackgroundTasks,
+    limit: int = Query(50, description="학습 종목 수"),
+):
+    """앙상블 모델 학습 (백그라운드 실행)"""
+    background_tasks.add_task(_train_task, limit)
+    return {
+        "message": f"학습이 백그라운드에서 시작되었습니다 ({limit}개 종목)",
+        "status": "training_started",
+    }
+
+
+@router.get("/backtest")
+async def get_backtest():
+    """모델 학습 메트릭 및 백테스트 결과 조회"""
+    ensemble = get_ensemble()
+    if not ensemble.is_trained:
+        raise HTTPException(status_code=503, detail="모델이 학습되지 않았습니다")
+
+    return {
+        "is_trained": True,
+        "metrics": ensemble.training_metrics,
+        "weights": {
+            "lgbm": ensemble.weights[0],
+            "lstm": ensemble.weights[1],
+            "rf": ensemble.weights[2],
+        },
+    }