添加批量解析

walk_forward.py

@@ -22,7 +22,6 @@ Walk-Forward 滚动验证
 from __future__ import annotations
 
 import argparse
-import glob
 import html
 import os
 import sys
@@ -38,27 +37,10 @@ OUT_DIR = "output"
 
 
 # --------------------------------------------------------------------------- #
-# 单段指标
+# 单段指标（从 mt5_report_parser 复用）
 # --------------------------------------------------------------------------- #
-def _seg_metrics(net: pd.Series) -> Dict[str, float]:
-    n = len(net)
-    if n == 0:
-        return {"n": 0, "net": 0.0, "pf": 0.0, "win": 0.0, "dd": 0.0, "exp": 0.0}
-    wins = net[net > 0]
-    losses = net[net <= 0]
-    gp = wins.sum()
-    gl = -losses.sum()
-    pf = gp / gl if gl > 0 else np.inf
-    eq = net.cumsum()
-    dd = float((eq - eq.cummax()).min())
-    return {
-        "n": int(n),
-        "net": float(net.sum()),
-        "pf": float(pf),
-        "win": float(len(wins) / n * 100),
-        "dd": dd,
-        "exp": float(net.mean()),
-    }
+
+from mt5_report_parser import compute_segment_metrics as _seg_metrics
 
 
 @dataclass
@@ -79,7 +61,7 @@ def walk_forward(
     step_days: Optional[int] = None,
 ) -> List[WFWindow]:
     """
-    滚动切窗。
+    滚动切窗（向量化边界查找）。
     trades: 含 open_time, net_profit 的 DataFrame
     is_days / oos_days: IS/OOS 窗口天数
     step_days: 滑动步长（默认 = oos_days，即不重叠的 OOS）
@@ -87,7 +69,12 @@ def walk_forward(
     if trades is None or trades.empty:
         return []
     t = trades.sort_values("open_time").reset_index(drop=True)
-    t["net_profit"] = t["net_profit"].astype(float)
+    times = t["open_time"].to_numpy()
+    nets = t["net_profit"].astype(float).to_numpy()
+    n_trades = len(nets)
+
+    # 预计算累积和（含前导0，cumsum[i] = sum of nets[:i]）
+    cumsum = np.concatenate([[0.0], np.cumsum(nets)])
 
     if step_days is None:
         step_days = oos_days
@@ -97,34 +84,66 @@ def walk_forward(
     total_span = (end - start).days
     win_span = is_days + oos_days
     if total_span < win_span:
-        return []  # 数据不足以切出一个窗口
+        return []
 
-    windows: List[WFWindow] = []
-    cur = start
+    # 用 numpy 二分查找窗口边界索引（所有窗口一次性确定）
+    cur = np.datetime64(start, "ns")
+    end_ns = np.datetime64(end, "ns")
+    bounds: List[Tuple[int, int, int, int]] = []  # (is_s, is_e, oos_s, oos_e) 为交易索引
     idx = 0
     while True:
-        is_s = cur
-        is_e = cur + pd.Timedelta(days=is_days)
+        is_e = cur + np.timedelta64(is_days, "D")
         oos_s = is_e
-        oos_e = oos_s + pd.Timedelta(days=oos_days)
-        if oos_e > end + pd.Timedelta(days=1):
+        oos_e = oos_s + np.timedelta64(oos_days, "D")
+        if oos_e > end_ns + np.timedelta64(1, "D"):
             break
-        is_mask = (t["open_time"] >= is_s) & (t["open_time"] < is_e)
-        oos_mask = (t["open_time"] >= oos_s) & (t["open_time"] < oos_e)
-        is_net = t.loc[is_mask, "net_profit"]
-        oos_net = t.loc[oos_mask, "net_profit"]
-        if len(is_net) == 0 or len(oos_net) == 0:
-            cur += pd.Timedelta(days=step_days)
-            continue
-        windows.append(WFWindow(
-            idx=idx,
-            is_start=is_s, is_end=is_e,
-            oos_start=oos_s, oos_end=oos_e,
-            is_metrics=_seg_metrics(is_net),
-            oos_metrics=_seg_metrics(oos_net),
-        ))
+        i_s = int(np.searchsorted(times, cur, side="left"))
+        i_e = int(np.searchsorted(times, is_e, side="left"))
+        o_s = int(np.searchsorted(times, oos_s, side="left"))
+        o_e = int(np.searchsorted(times, oos_e, side="left"))
+        bounds.append((i_s, i_e, o_s, o_e))
+        cur += np.timedelta64(step_days, "D")
         idx += 1
-        cur += pd.Timedelta(days=step_days)
+
+    windows: List[WFWindow] = []
+    for i, (i_s, i_e, o_s, o_e) in enumerate(bounds):
+        if i_e - i_s == 0 or o_e - o_s == 0:
+            continue
+        # IS 段指标（用累积和 O(1) 计算）
+        is_net_sum = float(cumsum[i_e] - cumsum[i_s])
+        is_n = i_e - i_s
+        is_seg = nets[i_s:i_e]
+        is_gp = float(np.sum(is_seg[is_seg > 0]))
+        is_gl = float(-np.sum(is_seg[is_seg <= 0]))
+        is_pf = is_gp / is_gl if is_gl > 0 else np.inf
+        is_cum = cumsum[i_s + 1:i_e + 1] - cumsum[i_s:i_e]
+        is_dd = float(np.min(is_cum - np.maximum.accumulate(is_cum))) if len(is_cum) else 0.0
+        is_win = float(np.sum(is_seg > 0) / is_n * 100) if is_n else 0.0
+        # OOS 段
+        oos_net_sum = float(cumsum[o_e] - cumsum[o_s])
+        oos_n = o_e - o_s
+        oos_seg = nets[o_s:o_e]
+        oos_gp = float(np.sum(oos_seg[oos_seg > 0]))
+        oos_gl = float(-np.sum(oos_seg[oos_seg <= 0]))
+        oos_pf = oos_gp / oos_gl if oos_gl > 0 else np.inf
+        oos_cum = cumsum[o_s + 1:o_e + 1] - cumsum[o_s:o_e]
+        oos_dd = float(np.min(oos_cum - np.maximum.accumulate(oos_cum))) if len(oos_cum) else 0.0
+        oos_win = float(np.sum(oos_seg > 0) / oos_n * 100) if oos_n else 0.0
+
+        windows.append(WFWindow(
+            idx=i,
+            is_start=times[i_s], is_end=times[min(i_e - 1, n_trades - 1)],
+            oos_start=times[o_s], oos_end=times[min(o_e - 1, n_trades - 1)],
+            is_metrics={
+                "n": is_n, "net": is_net_sum, "pf": is_pf,
+                "win": is_win, "dd": is_dd, "exp": is_net_sum / is_n if is_n else 0.0,
+            },
+            oos_metrics={
+                "n": oos_n, "net": oos_net_sum, "pf": oos_pf,
+                "win": oos_win, "dd": oos_dd, "exp": oos_net_sum / oos_n if oos_n else 0.0,
+            },
+        ))
+
     return windows