diff --git a/sanguo_factor/batch_eval.py b/sanguo_factor/batch_eval.py
index 13c8c9b..456eab9 100644
--- a/sanguo_factor/batch_eval.py
+++ b/sanguo_factor/batch_eval.py
@@ -144,9 +144,11 @@ def _eval_one(name: str, alpha_df: pl.DataFrame, cutoff_map: dict[str, _dt],
         # 优化:先 pivot 全表，再 pandas 切窗口+预热期广播掩码置 NaN（替代 per-factor hash join）
         # 语义等价性:join 版把「窗口外或 bar_idx<WARMUP」的行剔除；新版这些位置变 NaN
         # 下游 rank_corr_rows/metrics 所有指标以 NaN=缺失自动剔除，两者等价
-        Fw = res.pivot(index="datetime", on="vt_symbol", values="data").sort("datetime")
-        F = Fw.to_pandas().set_index("datetime")
+        # polars pivot 在 10M 行宽表化分钟级(py-spy实锤);唯一键下 pandas pivot=factorize+reshape 秒级
+        _pd = res.to_pandas()
+        F = _pd.pivot(index="datetime", columns="vt_symbol", values="data").sort_index()
         F.index = pd.to_datetime(F.index)
+        F.columns.name = None
         F = F.loc[start_dt:end_dt]  # 评估窗口切片
         # 预热期前置 NaN:每列用对应 cutoff_map[vt_symbol] 进行广播比较
         cut_arr = pd.Series([cutoff_map.get(s) for s in F.columns], index=F.columns).to_numpy()