codex_jxs_code/run_bb_midline_hier_search.py

"""
布林带中轨策略参数分层搜索（2020-2025）

说明:
- 全区间覆盖: period 1~1000, std 0.5~1000
- 分层搜索: 先粗扫全区间，再在候选周围细化，最终细化到 std=0.5 步长
- 使用 1m 触及方向过滤（先涨/先跌）时，按 period 复用触及方向以提速
"""

from __future__ import annotations

import argparse
import math
import os
import tempfile
import time
from collections import defaultdict
from concurrent.futures import ProcessPoolExecutor, as_completed
from pathlib import Path

import numpy as np
import pandas as pd

from strategy.bb_midline_backtest import BBMidlineConfig, run_bb_midline_backtest
from strategy.data_loader import get_1m_touch_direction, load_klines
from strategy.indicators import bollinger


G_DF: pd.DataFrame | None = None
G_DF_1M: pd.DataFrame | None = None
G_USE_1M: bool = True
G_STEP_MIN: int = 5


def frange(start: float, end: float, step: float) -> list[float]:
    out: list[float] = []
    x = float(start)
    while x <= end + 1e-9:
        out.append(round(x, 6))
        x += step
    return out


def build_grid(
    p_start: float,
    p_end: float,
    p_step: float,
    s_start: float,
    s_end: float,
    s_step: float,
) -> list[tuple[int, float]]:
    periods = sorted({max(1, min(1000, int(round(v)))) for v in frange(p_start, p_end, p_step)})
    stds = sorted({round(max(0.5, min(1000.0, v)), 2) for v in frange(s_start, s_end, s_step)})
    if 1000 not in periods:
        periods.append(1000)
    if 1000.0 not in stds:
        stds.append(1000.0)
    out = [(p, s) for p in periods for s in stds]
    return sorted(set(out))


def build_local_grid(
    centers: pd.DataFrame,
    p_window: int,
    p_step: int,
    s_window: float,
    s_step: float,
) -> list[tuple[int, float]]:
    out: set[tuple[int, float]] = set()
    for _, row in centers.iterrows():
        p0 = int(row["period"])
        s0 = float(row["std"])
        p_min = max(1, p0 - p_window)
        p_max = min(1000, p0 + p_window)
        s_min = max(0.5, s0 - s_window)
        s_max = min(1000.0, s0 + s_window)
        periods = sorted({max(1, min(1000, int(round(v)))) for v in frange(p_min, p_max, p_step)})
        stds = sorted({round(max(0.5, min(1000.0, v)), 2) for v in frange(s_min, s_max, s_step)})
        for p in periods:
            for s in stds:
                out.add((p, s))
    return sorted(out)


def score_row(ret_pct: float, sharpe: float, dd_pct: float, n_trades: int) -> float:
    # 偏向“收益稳定”: 收益和夏普加分，回撤和极少交易惩罚
    sparse_penalty = -5.0 if n_trades < 200 else 0.0
    return ret_pct + sharpe * 12.0 - dd_pct * 0.8 + sparse_penalty


def _init_worker(df_path: str, df_1m_path: str | None, use_1m: bool, step_min: int):
    global G_DF, G_DF_1M, G_USE_1M, G_STEP_MIN
    G_DF = pd.read_pickle(df_path)
    G_DF_1M = pd.read_pickle(df_1m_path) if (use_1m and df_1m_path) else None
    G_USE_1M = bool(use_1m)
    G_STEP_MIN = int(step_min)


def _eval_period_task(args: tuple[int, list[float]]) -> list[dict]:
    period, std_list = args
    assert G_DF is not None

    arr_touch_dir = None
    if G_USE_1M and G_DF_1M is not None:
        close = G_DF["close"].astype(float)
        bb_mid, _, _, _ = bollinger(close, period, 1.0)
        arr_touch_dir = get_1m_touch_direction(G_DF, G_DF_1M, bb_mid.values, kline_step_min=G_STEP_MIN)

    rows: list[dict] = []
    for std in std_list:
        cfg = BBMidlineConfig(
            bb_period=period,
            bb_std=float(std),
            initial_capital=200.0,
            margin_pct=0.01,
            leverage=100.0,
            cross_margin=True,
            fee_rate=0.0005,
            rebate_pct=0.90,
            rebate_hour_utc=0,
            fill_at_close=True,
            use_1m_touch_filter=G_USE_1M,
            kline_step_min=G_STEP_MIN,
        )
        result = run_bb_midline_backtest(
            G_DF,
            cfg,
            df_1m=G_DF_1M if G_USE_1M else None,
            arr_touch_dir_override=arr_touch_dir,
        )

        eq = result.equity_curve["equity"].dropna()
        if len(eq) == 0:
            final_eq = 0.0
            ret_pct = -100.0
            dd_u = -200.0
            dd_pct = 100.0
        else:
            final_eq = float(eq.iloc[-1])
            ret_pct = (final_eq - cfg.initial_capital) / cfg.initial_capital * 100.0
            dd_u = float((eq.astype(float) - eq.astype(float).cummax()).min())
            dd_pct = abs(dd_u) / cfg.initial_capital * 100.0

        n_trades = len(result.trades)
        win_rate = (
            sum(1 for t in result.trades if t.net_pnl > 0) / n_trades * 100.0
            if n_trades > 0
            else 0.0
        )
        pnl = result.daily_stats["pnl"].astype(float)
        sharpe = float(pnl.mean() / pnl.std()) * math.sqrt(365.0) if pnl.std() > 0 else 0.0
        stable_score = score_row(ret_pct, sharpe, dd_pct, n_trades)

        rows.append(
            {
                "period": period,
                "std": round(float(std), 2),
                "final_eq": final_eq,
                "ret_pct": ret_pct,
                "n_trades": n_trades,
                "win_rate": win_rate,
                "sharpe": sharpe,
                "max_dd_u": dd_u,
                "max_dd_pct": dd_pct,
                "stable_score": stable_score,
                "use_1m_filter": int(G_USE_1M),
            }
        )
    return rows


def evaluate_grid(
    params: list[tuple[int, float]],
    *,
    workers: int,
    df_path: str,
    df_1m_path: str | None,
    use_1m: bool,
    step_min: int,
    label: str,
) -> pd.DataFrame:
    by_period: dict[int, set[float]] = defaultdict(set)
    for p, s in params:
        by_period[int(p)].add(round(float(s), 2))

    tasks = [(p, sorted(stds)) for p, stds in sorted(by_period.items())]
    total_periods = len(tasks)
    total_combos = sum(len(stds) for _, stds in tasks)
    if total_combos == 0:
        return pd.DataFrame()

    print(f"[{label}] period组数={total_periods}, 参数组合={total_combos}, workers={workers}")
    start = time.time()
    rows: list[dict] = []
    done_periods = 0
    done_combos = 0

    with ProcessPoolExecutor(
        max_workers=workers,
        initializer=_init_worker,
        initargs=(df_path, df_1m_path, use_1m, step_min),
    ) as ex:
        future_map = {ex.submit(_eval_period_task, task): task for task in tasks}
        for fut in as_completed(future_map):
            period, stds = future_map[fut]
            res = fut.result()
            rows.extend(res)
            done_periods += 1
            done_combos += len(stds)
            if (
                done_periods == total_periods
                or done_periods % max(1, total_periods // 10) == 0
            ):
                elapsed = time.time() - start
                print(
                    f"[{label}] 进度 {done_combos}/{total_combos} 组合 "
                    f"({done_periods}/{total_periods} periods), {elapsed:.0f}s"
                )

    df = pd.DataFrame(rows)
    print(f"[{label}] 完成, 用时 {time.time() - start:.1f}s")
    return df


def summarize_yearly(eq: pd.Series, initial_capital: float = 200.0) -> pd.DataFrame:
    s = eq.dropna().copy()
    s.index = pd.to_datetime(s.index)
    out_rows: list[dict] = []
    prev = initial_capital
    for year in range(2020, 2026):
        sub = s[s.index.year == year]
        if len(sub) == 0:
            continue
        ye = float(sub.iloc[-1])
        ret = (ye - prev) / prev * 100.0 if prev > 0 else 0.0
        out_rows.append({"year": year, "year_end_equity": ye, "year_return_pct": ret})
        prev = ye
    return pd.DataFrame(out_rows)


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("-p", "--period", default="5m", choices=["5m", "15m", "30m"])
    parser.add_argument("--start", default="2020-01-01")
    parser.add_argument("--end", default="2026-01-01")
    parser.add_argument("-j", "--workers", type=int, default=max(1, (os.cpu_count() or 4) - 1))
    parser.add_argument("--no-1m", action="store_true", help="禁用 1m 方向过滤")
    args = parser.parse_args()

    use_1m = not args.no_1m
    step_min = int(args.period.replace("m", ""))

    out_dir = Path(__file__).resolve().parent / "strategy" / "results"
    out_dir.mkdir(parents=True, exist_ok=True)

    print(f"加载数据: {args.period} {args.start}~{args.end}")
    t0 = time.time()
    df = load_klines(args.period, args.start, args.end)
    df_1m = load_klines("1m", args.start, args.end) if use_1m else None
    print(
        f"  {args.period}: {len(df):,} 条"
        + (f", 1m: {len(df_1m):,} 条" if df_1m is not None else "")
        + f", {time.time()-t0:.1f}s\n"
    )

    with tempfile.NamedTemporaryFile(suffix=".pkl", delete=False) as f_df:
        df.to_pickle(f_df.name)
        df_path = f_df.name
    df_1m_path = None
    if df_1m is not None:
        with tempfile.NamedTemporaryFile(suffix=".pkl", delete=False) as f_1m:
            df_1m.to_pickle(f_1m.name)
            df_1m_path = f_1m.name

    try:
        evaluated: set[tuple[int, float]] = set()
        all_parts: list[pd.DataFrame] = []

        # Stage 1: 全区间粗扫
        stage1 = build_grid(1, 1000, 50, 0.5, 1000, 50)
        stage1 = [x for x in stage1 if x not in evaluated]
        df1 = evaluate_grid(
            stage1,
            workers=args.workers,
            df_path=df_path,
            df_1m_path=df_1m_path,
            use_1m=use_1m,
            step_min=step_min,
            label="stage1-global",
        )
        if not df1.empty:
            all_parts.append(df1)
            evaluated.update((int(r["period"]), float(r["std"])) for _, r in df1.iterrows())

        seed1 = (
            df1.sort_values("stable_score", ascending=False).head(6)
            if not df1.empty
            else pd.DataFrame(columns=["period", "std"])
        )

        # Stage 2: 候选周围中等步长细化
        stage2 = build_local_grid(seed1, p_window=25, p_step=5, s_window=50, s_step=10)
        stage2 = [x for x in stage2 if x not in evaluated]
        df2 = evaluate_grid(
            stage2,
            workers=args.workers,
            df_path=df_path,
            df_1m_path=df_1m_path,
            use_1m=use_1m,
            step_min=step_min,
            label="stage2-local",
        )
        if not df2.empty:
            all_parts.append(df2)
            evaluated.update((int(r["period"]), float(r["std"])) for _, r in df2.iterrows())

        pool2 = pd.concat([d for d in [df1, df2] if not d.empty], ignore_index=True)
        seed2 = (
            pool2.sort_values("stable_score", ascending=False).head(4)
            if len(pool2) > 0
            else pd.DataFrame(columns=["period", "std"])
        )

        # Stage 3: 候选周围更细化
        stage3 = build_local_grid(seed2, p_window=8, p_step=1, s_window=10, s_step=1)
        stage3 = [x for x in stage3 if x not in evaluated]
        df3 = evaluate_grid(
            stage3,
            workers=args.workers,
            df_path=df_path,
            df_1m_path=df_1m_path,
            use_1m=use_1m,
            step_min=step_min,
            label="stage3-fine",
        )
        if not df3.empty:
            all_parts.append(df3)
            evaluated.update((int(r["period"]), float(r["std"])) for _, r in df3.iterrows())

        pool3 = pd.concat([d for d in [df1, df2, df3] if not d.empty], ignore_index=True)
        seed3 = (
            pool3.sort_values("stable_score", ascending=False).head(2)
            if len(pool3) > 0
            else pd.DataFrame(columns=["period", "std"])
        )

        # Stage 4: 最终细化（std 0.5 步长）
        stage4 = build_local_grid(seed3, p_window=3, p_step=1, s_window=4, s_step=0.5)
        stage4 = [x for x in stage4 if x not in evaluated]
        df4 = evaluate_grid(
            stage4,
            workers=args.workers,
            df_path=df_path,
            df_1m_path=df_1m_path,
            use_1m=use_1m,
            step_min=step_min,
            label="stage4-final",
        )
        if not df4.empty:
            all_parts.append(df4)
            evaluated.update((int(r["period"]), float(r["std"])) for _, r in df4.iterrows())

        if not all_parts:
            raise RuntimeError("未得到任何评估结果")

        all_df = pd.concat(all_parts, ignore_index=True)
        all_df = all_df.drop_duplicates(subset=["period", "std"], keep="last")

        best_stable = all_df.sort_values("stable_score", ascending=False).iloc[0]
        best_return = all_df.sort_values("ret_pct", ascending=False).iloc[0]

        # 对最佳稳定参数再跑一次，导出逐年收益
        cfg = BBMidlineConfig(
            bb_period=int(best_stable["period"]),
            bb_std=float(best_stable["std"]),
            initial_capital=200.0,
            margin_pct=0.01,
            leverage=100.0,
            cross_margin=True,
            fee_rate=0.0005,
            rebate_pct=0.90,
            rebate_hour_utc=0,
            fill_at_close=True,
            use_1m_touch_filter=use_1m,
            kline_step_min=step_min,
        )
        final_res = run_bb_midline_backtest(df, cfg, df_1m=df_1m if use_1m else None)
        final_eq = final_res.equity_curve["equity"].dropna()
        yearly = summarize_yearly(final_eq, initial_capital=200.0)

        stamp = time.strftime("%Y%m%d_%H%M%S")
        all_path = out_dir / f"bb_midline_hier_search_{args.period}_{stamp}.csv"
        yearly_path = out_dir / f"bb_midline_hier_search_{args.period}_{stamp}_yearly.csv"
        all_df.sort_values("stable_score", ascending=False).to_csv(all_path, index=False)
        yearly.to_csv(yearly_path, index=False)

        print("\n" + "=" * 96)
        print("分层搜索完成")
        print(
            f"最佳稳定参数: period={int(best_stable['period'])}, std={float(best_stable['std']):.2f} | "
            f"final={best_stable['final_eq']:.4f}U | ret={best_stable['ret_pct']:+.2f}% | "
            f"dd={best_stable['max_dd_pct']:.2f}% | sharpe={best_stable['sharpe']:.3f} | "
            f"trades={int(best_stable['n_trades'])}"
        )
        print(
            f"最高收益参数: period={int(best_return['period'])}, std={float(best_return['std']):.2f} | "
            f"final={best_return['final_eq']:.4f}U | ret={best_return['ret_pct']:+.2f}% | "
            f"dd={best_return['max_dd_pct']:.2f}% | sharpe={best_return['sharpe']:.3f} | "
            f"trades={int(best_return['n_trades'])}"
        )
        print(f"结果文件: {all_path}")
        print(f"逐年文件: {yearly_path}")
        print("=" * 96)

    finally:
        Path(df_path).unlink(missing_ok=True)
        if df_1m_path:
            Path(df_1m_path).unlink(missing_ok=True)


if __name__ == "__main__":
    main()