DeepCoin/candle_features.py

"""
모든 봉 간격에 대해 BB 위치·캔들 형태(몸통/꼬리/높이) 특징을 계산하고
기준 타임라인(3분)에 맞춰 정렬합니다.
"""

from __future__ import annotations

import numpy as np
import pandas as pd

from config import ENTRY_INTERVAL
from strategy import prepare_entry_df

INTERVAL_LABELS: dict[int, str] = {
    3: "m3",
    10: "m10",
    15: "m15",
    30: "m30",
    60: "m60",
    240: "m240",
    1440: "d1",
}


def interval_prefix(interval: int) -> str:
    """컬럼 접두사 (예: m3, d1)."""
    return INTERVAL_LABELS.get(interval, f"m{interval}")


def compute_bar_features(df: pd.DataFrame) -> pd.DataFrame:
    """단일 봉 DataFrame에 위치·캔들 높이 특징을 추가합니다."""
    out = prepare_entry_df(df.copy())
    if len(out) < 2:
        return out

    o = out["Open"].astype(float)
    h = out["High"].astype(float)
    l = out["Low"].astype(float)
    c = out["Close"].astype(float)
    prev_c = c.shift(1)
    upper = out["Upper"].astype(float)
    lower = out["Lower"].astype(float)
    prev_upper = upper.shift(1)
    prev_lower = lower.shift(1)
    ma = out["MA"].astype(float)

    band = (upper - lower).replace(0, np.nan)
    out["bb_pos"] = ((c - lower) / band).clip(0, 1)
    out["bb_width_pct"] = (
        out["BB_Width"] if "BB_Width" in out.columns else (band / ma.replace(0, np.nan) * 100)
    )

    rng = (h - l).replace(0, np.nan)
    body = (c - o).abs()
    out["range_pct"] = (rng / c.replace(0, np.nan)) * 100
    out["body_ratio"] = (body / rng).fillna(0).clip(0, 1)
    out["upper_wick_ratio"] = ((h - np.maximum(o, c)) / rng).fillna(0).clip(0, 1)
    out["lower_wick_ratio"] = ((np.minimum(o, c) - l) / rng).fillna(0).clip(0, 1)
    out["ret_pct"] = ((c - prev_c) / prev_c.replace(0, np.nan)) * 100
    out["bullish"] = (c > o).astype(int)
    out["bearish"] = (c < o).astype(int)

    out["cross_up_lower"] = ((prev_c <= prev_lower) & (c > lower)).astype(int)
    out["cross_up_upper"] = ((prev_c < prev_upper) & (c >= upper)).astype(int)
    out["cross_down_lower"] = ((prev_c >= prev_lower) & (c < lower)).astype(int)
    out["below_lower"] = (c < lower).astype(int)
    out["above_upper"] = (c > upper).astype(int)
    out["inside_band"] = ((c >= lower) & (c <= upper)).astype(int)

    out["bb_pos_low"] = (out["bb_pos"] < 0.2).astype(int)
    out["bb_pos_high"] = (out["bb_pos"] > 0.8).astype(int)
    out["body_strong"] = (out["body_ratio"] > 0.55).astype(int)
    out["body_weak"] = (out["body_ratio"] < 0.25).astype(int)
    out["hammer"] = ((out["lower_wick_ratio"] > 0.45) & (out["body_ratio"] < 0.35)).astype(int)
    out["shooting_star"] = ((out["upper_wick_ratio"] > 0.45) & (out["body_ratio"] < 0.35)).astype(int)
    out["squeeze"] = (out["bb_width_pct"] < 0.8).astype(int)

    return out


FEATURE_BOOL_COLS: tuple[str, ...] = (
    "cross_up_lower",
    "cross_up_upper",
    "cross_down_lower",
    "below_lower",
    "above_upper",
    "inside_band",
    "bb_pos_low",
    "bb_pos_high",
    "body_strong",
    "body_weak",
    "hammer",
    "shooting_star",
    "squeeze",
    "bullish",
    "bearish",
)


def _merge_interval_features(
    master_index: pd.DatetimeIndex,
    feat: pd.DataFrame,
    prefix: str,
) -> pd.DataFrame:
    """master_index 길이와 동일한 간격 특징만 반환."""
    pick = [c for c in FEATURE_BOOL_COLS if c in feat.columns]
    extra = [c for c in ("bb_pos", "body_ratio", "lower_wick_ratio", "ret_pct") if c in feat.columns]
    sub = feat[pick + extra].copy()
    sub.columns = [f"{prefix}_{c}" for c in sub.columns]

    left = pd.DataFrame({"ts": master_index})
    right = sub.reset_index()
    time_col = right.columns[0]
    right = right.rename(columns={time_col: "ts"})

    merged = pd.merge_asof(
        left.sort_values("ts"),
        right.sort_values("ts"),
        on="ts",
        direction="backward",
    )
    merged.index = master_index
    return merged.drop(columns=["ts"])


def build_master_feature_matrix(frames: dict[int, pd.DataFrame]) -> pd.DataFrame:
    """3분 타임라인에 모든 봉의 위치·캔들 특징을 붙인 행렬 (인덱스 유일)."""
    entry = frames.get(ENTRY_INTERVAL)
    if entry is None or entry.empty:
        raise ValueError("ENTRY_INTERVAL 데이터가 없습니다.")

    entry_feat = compute_bar_features(entry)
    entry_feat = entry_feat[~entry_feat.index.duplicated(keep="last")].sort_index()

    p3 = interval_prefix(ENTRY_INTERVAL)
    ohlc = ["Open", "High", "Low", "Close", "Volume", "Upper", "Lower", "MA"]
    master = entry_feat[[c for c in ohlc if c in entry_feat.columns]].copy()

    for col in FEATURE_BOOL_COLS:
        if col in entry_feat.columns:
            master[f"{p3}_{col}"] = entry_feat[col]
    for col in ("bb_pos", "body_ratio", "lower_wick_ratio", "ret_pct", "bb_width_pct"):
        if col in entry_feat.columns:
            master[f"{p3}_{col}"] = entry_feat[col]

    parts = [master]
    for interval, df in sorted(frames.items()):
        if interval == ENTRY_INTERVAL or df is None or df.empty:
            continue
        feat = compute_bar_features(df)
        feat = feat[~feat.index.duplicated(keep="last")].sort_index()
        prefix = interval_prefix(interval)
        parts.append(_merge_interval_features(master.index, feat, prefix))

    out = pd.concat(parts, axis=1)
    return out.loc[:, ~out.columns.duplicated()]