From b1ca6ab5d31b6775a1244946e957305276dabca5 Mon Sep 17 00:00:00 2001
From: tkrmagid <tkrmagid@git.tkrmagid.kr>
Date: Wed, 20 May 2026 15:59:14 +0900
Subject: [PATCH] =?UTF-8?q?feat(phase-3):=20Chronos=20zero-shot=20?=
 =?UTF-8?q?=EC=98=88=EC=B8=A1=20+=20=ED=94=BC=EC=B2=98=20=EB=B9=8C?=
 =?UTF-8?q?=EB=8D=94?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

- backend/app/models/chronos.py: amazon/chronos-t5-small (env CHRONOS_MODEL
  override 가능). lazy singleton, cuda + bf16 자동, q10/median/q90 + raw
  samples 반환 (앙상블 가중평균용).
- backend/app/models/features.py: 종목별 학습/추론 피처 DataFrame.
  OHLCV + TA(rsi/macd/atr/bb/sma/ema/vol_z) + 외인기관거래대금 + macro
  (kospi/kosdaq/usdkrw/us10y + r1) + sentiment(v_sentiment_daily, 3d rolling).
  학습용은 with_targets=True 로 y_close_h{1,3,5}, y_ret_h*, y_dir_h*
  (±0.3% flat band) 추가.
- pyproject.toml: chronos-forecasting 1.4.1, accelerate 0.30.1, joblib 1.4.2.

이 단계까지는 코드만. 실제 모델 다운로드는 첫 ping/predict 호출 시점에.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
---
 backend/app/models/chronos.py  | 118 +++++++++++++++++
 backend/app/models/features.py | 223 +++++++++++++++++++++++++++++++++
 backend/pyproject.toml         |   3 +
 3 files changed, 344 insertions(+)
 create mode 100644 backend/app/models/chronos.py
 create mode 100644 backend/app/models/features.py

diff --git a/backend/app/models/chronos.py b/backend/app/models/chronos.py
new file mode 100644
index 0000000..7f660f6
--- /dev/null
+++ b/backend/app/models/chronos.py
@@ -0,0 +1,118 @@
+"""Chronos zero-shot 시계열 예측 어댑터.
+
+모델: amazon/chronos-t5-small (46M, 빠르고 RTX 3070 Ti 에 충분히 들어감).
+환경변수 CHRONOS_MODEL 로 base/large 로 바꿀 수 있음.
+
+입력: 종가 시계열 (list[float], 최소 32 step).
+출력: horizon 일 quantile forecast (q10/median/q90).
+
+lazy singleton 으로 첫 호출 시 모델 로드. 디바이스는 settings.model_device 따라.
+"""
+from __future__ import annotations
+
+import logging
+import os
+import threading
+from dataclasses import dataclass
+
+from app.config import settings
+
+logger = logging.getLogger(__name__)
+
+MODEL_NAME = os.environ.get("CHRONOS_MODEL", "amazon/chronos-t5-small")
+
+_lock = threading.Lock()
+_state: dict[str, object] = {"loaded": False, "pipe": None, "device": None}
+
+
+@dataclass
+class ChronosForecast:
+    horizon: int
+    median: list[float]
+    q10: list[float]
+    q90: list[float]
+    samples: list[list[float]]  # raw samples for ensemble downstream
+
+
+def _resolve_device() -> str:
+    import torch  # lazy
+
+    pref = (settings.model_device or "auto").lower()
+    if pref == "cuda":
+        return "cuda" if torch.cuda.is_available() else "cpu"
+    if pref == "cpu":
+        return "cpu"
+    return "cuda" if torch.cuda.is_available() else "cpu"
+
+
+def _load() -> None:
+    global _state
+    with _lock:
+        if _state["loaded"]:
+            return
+        import torch
+        from chronos import ChronosPipeline
+
+        token = settings.huggingface_token or None
+        if token:
+            os.environ.setdefault("HUGGINGFACE_HUB_TOKEN", token)
+            os.environ.setdefault("HF_TOKEN", token)
+
+        device = _resolve_device()
+        # bf16 은 RTX 30xx 이상에서 지원. cpu 에선 fp32.
+        dtype = torch.bfloat16 if device == "cuda" else torch.float32
+        logger.info("loading Chronos %s on %s (dtype=%s)", MODEL_NAME, device, dtype)
+        pipe = ChronosPipeline.from_pretrained(
+            MODEL_NAME,
+            device_map=device,
+            torch_dtype=dtype,
+        )
+        _state.update({"loaded": True, "pipe": pipe, "device": device})
+
+
+def forecast(
+    series: list[float],
+    *,
+    horizon: int = 5,
+    num_samples: int = 30,
+) -> ChronosForecast:
+    """series 의 마지막 시점 이후 horizon 일 예측.
+
+    series 는 일봉 종가. 최소 32개 권장 (그보다 짧으면 Chronos 분위 안정성 떨어짐).
+    """
+    if len(series) < 32:
+        raise ValueError(
+            f"series too short ({len(series)}) for Chronos forecast (need >=32)"
+        )
+    _load()
+    import numpy as np
+    import torch
+
+    pipe = _state["pipe"]
+    context = torch.tensor([float(x) for x in series], dtype=torch.float32)
+    with torch.no_grad():
+        samples = pipe.predict(
+            context=context,
+            prediction_length=horizon,
+            num_samples=num_samples,
+        )
+    # samples: (1, num_samples, prediction_length)
+    arr = samples[0].cpu().float().numpy()
+    q10 = np.quantile(arr, 0.10, axis=0).tolist()
+    median = np.quantile(arr, 0.50, axis=0).tolist()
+    q90 = np.quantile(arr, 0.90, axis=0).tolist()
+    return ChronosForecast(
+        horizon=horizon,
+        median=[float(x) for x in median],
+        q10=[float(x) for x in q10],
+        q90=[float(x) for x in q90],
+        samples=[[float(x) for x in row] for row in arr.tolist()],
+    )
+
+
+def ping() -> dict[str, object]:
+    try:
+        _load()
+        return {"status": "ok", "model": MODEL_NAME, "device": _state["device"]}
+    except Exception as exc:  # noqa: BLE001
+        return {"status": "failed", "model": MODEL_NAME, "error": str(exc)}
diff --git a/backend/app/models/features.py b/backend/app/models/features.py
new file mode 100644
index 0000000..e000b05
--- /dev/null
+++ b/backend/app/models/features.py
@@ -0,0 +1,223 @@
+"""모델 학습/추론용 피처 빌더.
+
+종목 1개 + 룩백 기간을 받아 (date 단위) DataFrame 반환:
+  - OHLCV
+  - returns r1
+  - TA: rsi14, macd, macd_signal, atr14, bb_pct, sma20, ema12, vol_z20
+  - trading_value: foreign_net, institution_net, individual_net (정규화 X, scale 그대로)
+  - macro 정렬: kospi, kosdaq, usdkrw, us10y, kospi_r1, usdkrw_r1
+  - sentiment (v_sentiment_daily): mean_score, weighted_score, n_articles,
+    pos_minus_neg = pos_ratio - neg_ratio. 3일 롤링 mean 도 추가.
+
+학습 타깃 (build_features 에서만 생성):
+  - y_close_h{1,3,5}: close.shift(-H)
+  - y_ret_h{1,3,5}:   y_close_h / close - 1
+  - y_dir_h{1,3,5}:   sign(y_ret_h) (1=up, -1=down, 0=flat ±0.3% 이내)
+
+inference 용 build_features 는 dropna 안 함. 학습용 build_training_frame 은 dropna.
+"""
+from __future__ import annotations
+
+import logging
+from dataclasses import dataclass
+from datetime import date, timedelta
+
+import numpy as np
+import pandas as pd
+from sqlalchemy import text
+
+from app.db.connection import get_engine
+
+logger = logging.getLogger(__name__)
+
+FLAT_BAND = 0.003  # ±0.3% 이내는 flat
+HORIZONS_DEFAULT = (1, 3, 5)
+
+
+@dataclass
+class FeatureFrame:
+    code: str
+    df: pd.DataFrame
+    target_horizons: tuple[int, ...]
+
+
+def _load_ohlcv(code: str, start: date, end: date) -> pd.DataFrame:
+    eng = get_engine()
+    sql = text(
+        """
+        SELECT date, open, high, low, close, volume
+          FROM ohlcv_daily
+         WHERE code = :code AND date BETWEEN :s AND :e
+         ORDER BY date
+        """
+    )
+    with eng.connect() as conn:
+        rows = conn.execute(sql, {"code": code, "s": start, "e": end}).all()
+    if not rows:
+        return pd.DataFrame(columns=["date", "open", "high", "low", "close", "volume"])
+    df = pd.DataFrame(rows, columns=["date", "open", "high", "low", "close", "volume"])
+    df["date"] = pd.to_datetime(df["date"]).dt.date
+    return df
+
+
+def _load_trading(code: str, start: date, end: date) -> pd.DataFrame:
+    eng = get_engine()
+    sql = text(
+        """
+        SELECT date, foreign_net, institution_net, individual_net
+          FROM trading_value_daily
+         WHERE code = :code AND date BETWEEN :s AND :e
+         ORDER BY date
+        """
+    )
+    with eng.connect() as conn:
+        rows = conn.execute(sql, {"code": code, "s": start, "e": end}).all()
+    if not rows:
+        return pd.DataFrame(columns=["date", "foreign_net", "institution_net", "individual_net"])
+    df = pd.DataFrame(rows, columns=["date", "foreign_net", "institution_net", "individual_net"])
+    df["date"] = pd.to_datetime(df["date"]).dt.date
+    return df
+
+
+def _load_macro(start: date, end: date) -> pd.DataFrame:
+    eng = get_engine()
+    sql = text(
+        "SELECT date, key, value FROM macro_daily "
+        "WHERE date BETWEEN :s AND :e ORDER BY date"
+    )
+    with eng.connect() as conn:
+        rows = conn.execute(sql, {"s": start, "e": end}).all()
+    if not rows:
+        return pd.DataFrame(columns=["date"])
+    df = pd.DataFrame(rows, columns=["date", "key", "value"])
+    pivot = df.pivot_table(index="date", columns="key", values="value", aggfunc="last").reset_index()
+    pivot["date"] = pd.to_datetime(pivot["date"]).dt.date
+    pivot.columns.name = None
+    return pivot
+
+
+def _load_sentiment(code: str, start: date, end: date) -> pd.DataFrame:
+    eng = get_engine()
+    sql = text(
+        """
+        SELECT date, n_articles, mean_score, pos_ratio, neg_ratio,
+               weighted_score
+          FROM v_sentiment_daily
+         WHERE code = :code AND date BETWEEN :s AND :e
+         ORDER BY date
+        """
+    )
+    with eng.connect() as conn:
+        rows = conn.execute(sql, {"code": code, "s": start, "e": end}).all()
+    cols = ["date", "n_articles", "mean_score", "pos_ratio", "neg_ratio", "weighted_score"]
+    if not rows:
+        return pd.DataFrame(columns=cols)
+    df = pd.DataFrame(rows, columns=cols)
+    df["date"] = pd.to_datetime(df["date"]).dt.date
+    df["pos_minus_neg"] = df["pos_ratio"].fillna(0) - df["neg_ratio"].fillna(0)
+    return df
+
+
+def _add_ta(df: pd.DataFrame) -> pd.DataFrame:
+    """ta 패키지로 기술 지표 추가."""
+    from ta.momentum import RSIIndicator
+    from ta.trend import EMAIndicator, MACD, SMAIndicator
+    from ta.volatility import AverageTrueRange, BollingerBands
+
+    close = df["close"].astype(float)
+    high = df["high"].astype(float)
+    low = df["low"].astype(float)
+    vol = df["volume"].astype(float)
+
+    df["r1"] = close.pct_change()
+    df["rsi14"] = RSIIndicator(close=close, window=14, fillna=False).rsi()
+    macd = MACD(close=close, window_slow=26, window_fast=12, window_sign=9, fillna=False)
+    df["macd"] = macd.macd()
+    df["macd_signal"] = macd.macd_signal()
+    df["atr14"] = AverageTrueRange(high=high, low=low, close=close, window=14, fillna=False).average_true_range()
+    bb = BollingerBands(close=close, window=20, window_dev=2, fillna=False)
+    df["bb_pct"] = bb.bollinger_pband()
+    df["sma20"] = SMAIndicator(close=close, window=20, fillna=False).sma_indicator()
+    df["ema12"] = EMAIndicator(close=close, window=12, fillna=False).ema_indicator()
+    vol_mean = vol.rolling(20).mean()
+    vol_std = vol.rolling(20).std().replace(0, np.nan)
+    df["vol_z20"] = (vol - vol_mean) / vol_std
+    return df
+
+
+def _add_targets(df: pd.DataFrame, horizons: tuple[int, ...]) -> pd.DataFrame:
+    close = df["close"].astype(float)
+    for h in horizons:
+        df[f"y_close_h{h}"] = close.shift(-h)
+        df[f"y_ret_h{h}"] = df[f"y_close_h{h}"] / close - 1.0
+        df[f"y_dir_h{h}"] = np.where(
+            df[f"y_ret_h{h}"] > FLAT_BAND, 1,
+            np.where(df[f"y_ret_h{h}"] < -FLAT_BAND, -1, 0),
+        )
+    return df
+
+
+def build_features(
+    code: str,
+    *,
+    lookback_days: int = 365 * 2,
+    end_date: date | None = None,
+    horizons: tuple[int, ...] = HORIZONS_DEFAULT,
+    with_targets: bool = False,
+) -> FeatureFrame:
+    """code 1개 종목의 피처 DataFrame 생성.
+
+    inference: with_targets=False 로 호출 → 최신 row 의 피처만 LGBM/Chronos 에 투입.
+    training : with_targets=True 로 호출 → tail H 행은 타깃 NaN → dropna 로 제거.
+    """
+    end = end_date or date.today()
+    start = end - timedelta(days=lookback_days)
+
+    ohlcv = _load_ohlcv(code, start, end)
+    if ohlcv.empty:
+        return FeatureFrame(code=code, df=ohlcv, target_horizons=horizons)
+
+    df = ohlcv.copy().sort_values("date").reset_index(drop=True)
+
+    df = _add_ta(df)
+
+    trading = _load_trading(code, start, end)
+    if not trading.empty:
+        df = df.merge(trading, on="date", how="left")
+    else:
+        for col in ("foreign_net", "institution_net", "individual_net"):
+            df[col] = np.nan
+
+    macro = _load_macro(start, end)
+    if not macro.empty:
+        df = df.merge(macro, on="date", how="left")
+        for k in ("kospi", "kosdaq", "usdkrw", "us10y"):
+            if k in df.columns:
+                df[f"{k}_r1"] = df[k].pct_change()
+
+    sentiment = _load_sentiment(code, start, end)
+    if not sentiment.empty:
+        df = df.merge(sentiment, on="date", how="left")
+        # 3일 롤링 평균
+        for col in ("mean_score", "weighted_score", "pos_minus_neg", "n_articles"):
+            if col in df.columns:
+                df[f"{col}_3d"] = df[col].rolling(3, min_periods=1).mean()
+    else:
+        for col in ("n_articles", "mean_score", "pos_ratio", "neg_ratio",
+                    "weighted_score", "pos_minus_neg"):
+            df[col] = np.nan
+
+    if with_targets:
+        df = _add_targets(df, horizons)
+
+    return FeatureFrame(code=code, df=df, target_horizons=horizons)
+
+
+def feature_columns(df: pd.DataFrame) -> list[str]:
+    """LGBM 학습/추론용 피처 컬럼 목록. date / OHLCV / y_* 제외."""
+    drop = {"date", "open", "high", "low", "close", "volume"}
+    cols = [
+        c for c in df.columns
+        if c not in drop and not c.startswith("y_")
+    ]
+    return cols
diff --git a/backend/pyproject.toml b/backend/pyproject.toml
index e4c7a35..8a97bca 100644
--- a/backend/pyproject.toml
+++ b/backend/pyproject.toml
@@ -31,9 +31,12 @@ dependencies = [
     "transformers==4.41.2",
     "tokenizers==0.19.1",
     "sentencepiece==0.2.0",
+    "accelerate==0.30.1",
+    "chronos-forecasting==1.4.1",
     "scikit-learn==1.5.0",
     "lightgbm==4.3.0",
     "ta==0.11.0",
+    "joblib==1.4.2",
 
     # scheduler
     "apscheduler==3.10.4",