"""按**实盘口径**重放 B4，并测三道过滤能否刷掉那批亏钱的额外信号。 step63 的结论是一好一坏： 好 时点干净 —— 100% 召回、100% 准时、零滞后（一类是 0% 准时、+15 根） 坏 回放多出 592 个全量口径没有的信号，PF 0.46 / t −5.03，混合后 0.64 < 1 但 step63 有两个口径问题，本脚本一并修掉： ① 窗口不对。回测用全量 45000 根一次算完，step63 用 2 万涨到 4 万根的增长窗口， **而实盘用 2001 根滚动窗口、每 500 根 init_stream 拉回**（`shadow_signal.py` 的 MAX_GROW）。三种口径的中枢结构都不一样。这也是 HANDOFF 里挂着的 「回测用全量历史建中枢、实盘用 2000 根窗口」那条待办。 顺带：窗口封顶后单步成本恒定，不再是 step63 那个平方级（143ms@2万根 -> 292ms@4万根），所以本脚本快得多。 ② 没测过滤。step63 跑的是裸信号，而实盘有三道滤网。ATR 门控已单独测过—— 它刷掉 7.9% 的额外信号却刷掉 10.7% 的好信号，PF 纹丝不动。剩下两道要测。 **一处刻意的简化，方向是保守的**：大级别分型时间线用全量历史算（真实盘的 HTF 也会重画）。这等于**给同向过滤器开了未来函数的后门**。若连这样都刷不掉额外信号， 结论只会更强。 """ from __future__ import annotations import argparse import sys import warnings from concurrent.futures import ProcessPoolExecutor, as_completed from pathlib import Path import numpy as np import pandas as pd warnings.filterwarnings("ignore") HERE = Path(__file__).resolve().parent sys.path.insert(0, str(HERE)) sys.path.insert(0, str(HERE.parent)) OUT = HERE / "out" / "step68_live_window.feather" SL, SCALE_AT, RUNNER, RSTOP, MAXB = 2.0, 3.0, 8.0, 2.0, 48 WIN, MAX_GROW, GATE_BP = 2001, 500, 8.0 def _ladder(zones: pd.DataFrame) -> pd.DataFrame: z = zones.copy() pg, pdn = z["zg"].shift(), z["zd"].shift() z["z_above"], z["z_below"] = z["zd"] > pg, z["zg"] < pdn z["zone_i"] = np.arange(len(z)) return z def replay(sym: str, ltf: str, htf: str, rows: int, steps: int) -> pd.DataFrame | None: from chanlun import TF_DF from chanlun.analysis.fast_bsp import ( ensure_timestamp, find_fast_bsp3, zones_from_zs_list, ) from lib.data import fetch_ohlcv from lib.fx_signal import extract_fx_signals, signals_to_frame from lib.nested_bsp import attach_htf_context, htf_fx_timeline try: df = fetch_ohlcv(f"{sym}/USDT:USDT", ltf, rows) if df is None or len(df) < WIN + steps + 100: return None df = df.iloc[-(WIN + steps):].reset_index(drop=True) full = TF_DF(df, 1, ltf) cdf = ensure_timestamp(full.dataframe) zs_full = full.cal_bi_zs_list_pure(full.bi_list) if not zs_full: return None sig_full = find_fast_bsp3(cdf, zones_from_zs_list(zs_full, cdf)) full_keys = {(int(r.entry_idx), int(r.direction)) for r in sig_full.itertuples()} if not sig_full.empty \ else set() # 大级别分型时间线：全量算（见模块 docstring 的「刻意简化」） dh = fetch_ohlcv(f"{sym}/USDT:USDT", htf, rows) tl = None if dh is not None and len(dh) > 500: ch = TF_DF(dh, 1, htf) tl = htf_fx_timeline( signals_to_frame(extract_fx_signals(ch, ch.dataframe)), ch.dataframe) rec: dict[tuple, dict] = {} chan = None anchor = 0 for i in range(WIN, len(df)): # 实盘的窗口纪律：2001 根起，长过 MAX_GROW 就 init_stream 拉回 if chan is None or (i - anchor) >= MAX_GROW: w = df.iloc[i - WIN + 1:i + 1].copy() chan = TF_DF(w, 1, ltf) chan.init_stream(w, 1, ltf) anchor = i else: chan.append_bar(df.iloc[i]) try: zl = chan.cal_bi_zs_list_pure(chan.bi_list) if not zl: continue sub = ensure_timestamp(chan.dataframe) z = _ladder(zones_from_zs_list(zl, sub)) if z.empty: continue s = find_fast_bsp3(sub, z) if s is None or s.empty: continue last = len(sub) - 1 s = s[s["entry_idx"].astype(int) == last] # 实盘只做当根 if s.empty: continue if "zone_i" in s.columns: s = s.merge(z[["zone_i", "z_above", "z_below"]], on="zone_i", how="left") if tl is not None: s = attach_htf_context(s, sub, tl, "h1") except Exception: # noqa: BLE001 continue for r in s.itertuples(): k = (i, int(r.direction)) if k in rec: continue push = getattr(r, "z_above" if r.direction == 1 else "z_below", None) ag = getattr(r, "h1_agree", 0) rec[k] = { "sym": sym, "entry_idx": i, "direction": int(r.direction), "in_full": (i, int(r.direction)) in full_keys, "ladder_ok": int(bool(pd.notna(push) and bool(push))), "h1_agree": int(ag) if pd.notna(ag) else 0, } if not rec: return None r = pd.DataFrame(list(rec.values())) from lib.exit_model import cfg_name, walk_exits atr = cdf["atr"].to_numpy(float) cl = cdf["close"].to_numpy(float) r = r[(r.entry_idx < len(cdf) - 2) & np.isfinite(atr[r.entry_idx.values]) & (atr[r.entry_idx.values] > 0)].reset_index(drop=True) if r.empty: return None res = walk_exits(cdf, pd.DataFrame({ "entry_idx": r.entry_idx.values, "direction": r.direction.values}), [SL], [RUNNER], [MAXB], scale_at=SCALE_AT, runners=(RUNNER,), runner_stops=(RSTOP,)) cfg = cfg_name(SL, RUNNER, MAXB, RSTOP) if len(res) != len(r): return None for c in ("g", "r", "c"): r[c] = res[f"{cfg}_{c}"].to_numpy() r["atr_pct"] = atr[r.entry_idx.values] / cl[r.entry_idx.values] r["gate_ok"] = (r.atr_pct * 1e4 >= GATE_BP).astype(int) r["pass_all"] = ((r.h1_agree == 1) & (r.ladder_ok == 1) & (r.gate_ok == 1)).astype(int) return r except Exception as e: # noqa: BLE001 print(f" {sym} 失败: {type(e).__name__}: {e}", flush=True) return None def perf(g: pd.DataFrame) -> dict | None: from lib.exit_model import fee_of, taker_notional if len(g) < 20: return None net = g.g.values - fee_of(g.r.values, g.c.values) gR = g.g.values / (SL * g.atr_pct.values) tn = taker_notional(g.r.values, g.c.values) w, o = net[net > 0].sum(), -net[net <= 0].sum() return { "笔数": len(g), "胜率": f"{(net > 0).mean()*100:.1f}%", "PF": round(w / o, 2) if o > 0 else np.inf, "余量bp": round(net.mean() / tn.mean() * 1e4, 2), "t值": round(gR.mean() / (gR.std(ddof=1) / np.sqrt(len(g))), 2), } def report(d: pd.DataFrame) -> None: print("\n" + "=" * 92) print("【一】实盘窗口下还有多少额外信号") print("=" * 92) print(f"回放信号 {len(d)} · 其中全量口径也有 {int(d.in_full.sum())} · " f"**额外 {int((~d.in_full).sum())}**") print(f"（step63 的增长窗口口径：197 / 592）") print("\n" + "=" * 92) print("【二】三道过滤能不能刷掉额外信号 —— 这决定实盘是否在亏钱") print("=" * 92) rows = [] for nm, m in [("① 无过滤", None), ("② 仅同向", d.h1_agree == 1), ("③ 仅阶梯", d.ladder_ok == 1), ("④ 仅ATR门控", d.gate_ok == 1), ("⑤ 三道全开(实盘口径)", d.pass_all == 1)]: x = d if m is None else d[m] if x.empty: continue ex, bt = x[~x.in_full], x[x.in_full] row = {"过滤": nm, "留下": len(x), "额外占比": f"{(~x.in_full).mean()*100:.0f}%"} for lab, g in [("全部", x), ("额外", ex), ("回测口径", bt)]: s = perf(g) row[f"{lab}PF"] = "—" if s is None else s["PF"] row[f"{lab}n"] = len(g) rows.append(row) print(pd.DataFrame(rows).to_string(index=False)) print("\n" + "=" * 92) print("【三】实盘口径（三道全开）的完整表现") print("=" * 92) rows = [] for nm, g in [("实盘会做的全部", d[d.pass_all == 1]), (" 其中额外的", d[(d.pass_all == 1) & ~d.in_full]), (" 其中回测也有的", d[(d.pass_all == 1) & d.in_full])]: s = perf(g) if s: rows.append({"分组": nm, **s}) print(pd.DataFrame(rows).to_string(index=False)) print(""" 判读：「实盘会做的全部」PF > 1 -> 实盘安全，额外信号被滤网挡住了 PF < 1 -> **实盘在做一批回测里不存在、且亏钱的信号**，要立刻处理""") def main() -> None: ap = argparse.ArgumentParser() ap.add_argument("--symbols", default="BTC,ETH,SOL,LINK,DOGE") ap.add_argument("--ltf", default="5m") ap.add_argument("--htf", default="30m") ap.add_argument("--rows", type=int, default=45_000) ap.add_argument("--steps", type=int, default=20_000) ap.add_argument("--workers", type=int, default=5) ap.add_argument("--reuse", action="store_true") args = ap.parse_args() if args.reuse and OUT.exists(): report(pd.read_feather(OUT)) return syms = [s.strip() for s in args.symbols.split(",")] print(f"[实盘口径回放] {len(syms)} 币 × {args.steps} 根 · " f"{WIN} 根滚动窗口 / 每 {MAX_GROW} 根重建\n", flush=True) parts = [] with ProcessPoolExecutor(max_workers=args.workers) as ex: fut = {ex.submit(replay, s, args.ltf, args.htf, args.rows, args.steps): s for s in syms} for i, f in enumerate(as_completed(fut), 1): r = f.result() print(f" [{i}/{len(syms)}] {fut[f]} " f"{0 if r is None else len(r)} 个信号", flush=True) if r is not None: parts.append(r) if not parts: print("无结果") return d = pd.concat(parts, ignore_index=True) OUT.parent.mkdir(exist_ok=True) d.to_feather(OUT) report(d) if __name__ == "__main__": main()