Improve Trader entry quality training diagnostics

training/trader_training/entry_feature_screen.py

@@ -24,9 +24,11 @@ def screen_entry_features(args: Any) -> None:
 
     min_bucket_rows = int(args.min_bucket_rows or 300)
     rows: list[dict[str, Any]] = []
+    edge_source_by_side: dict[str, str] = {}
     for side in ("LONG", "SHORT"):
         target_col = "long_entry_target" if side == "LONG" else "short_entry_target"
-        edge_col = "long_expected_net_edge_bps" if side == "LONG" else "short_expected_net_edge_bps"
+        edge_col = _screen_edge_column(dataset, side)
+        edge_source_by_side[side] = edge_col
         baselines = _split_baselines(dataset, target_col, edge_col)
         for feature in FEATURE_ORDER:
             rows.extend(_feature_rows(dataset, feature, side, target_col, edge_col, baselines))
@@ -43,6 +45,7 @@ def screen_entry_features(args: Any) -> None:
         "bucket_metric_count": int(len(bucket_metrics)),
         "candidate_count": int(len(candidates)),
         "min_bucket_rows": min_bucket_rows,
+        "edge_source_by_side": edge_source_by_side,
         "selection_rule": "bucket boundaries are learned on fit_inner; candidate is picked by tune_inner and checked on validation_locked/latest_stress",
     }
     write_json(root / "diagnostics" / "entry_feature_screen_result.json", result)
@@ -59,6 +62,14 @@ def screen_entry_features(args: Any) -> None:
     )
 
 
+def _screen_edge_column(dataset: pd.DataFrame, side: str) -> str:
+    prefix = "long" if side == "LONG" else "short"
+    actual_col = f"{prefix}_actual_plan_net_edge_bps"
+    if actual_col in dataset.columns:
+        return actual_col
+    return f"{prefix}_expected_net_edge_bps"
+
+
 def _split_baselines(dataset: pd.DataFrame, target_col: str, edge_col: str) -> dict[str, dict[str, float]]:
     baselines: dict[str, dict[str, float]] = {}
     for split_id in ALL_SPLITS:
@@ -225,6 +236,7 @@ def _markdown_report(result: dict[str, Any], candidates: pd.DataFrame) -> str:
         "",
         "这份报告只回答一个问题：历史数据里，单个特征的某些区间有没有稳定变好。",
         "",
+        "- 如果数据里有真实出场净收益，本报告用真实出场净收益；没有时才退回训练收益标签。",
         "- `tune_inner` 用来挑候选区间。",
         "- `validation_locked` 和 `latest_stress` 用来检查这个区间是不是出了训练样本也还能站住。",
         "- `stable_positive_edge=true` 代表这个区间在三个检查集里的平均净收益都大于 0。",
@@ -237,6 +249,8 @@ def _markdown_report(result: dict[str, Any], candidates: pd.DataFrame) -> str:
         f"- 分桶明细数: `{result['bucket_metric_count']}`",
         f"- 候选数: `{result['candidate_count']}`",
         f"- 最小分桶行数: `{result['min_bucket_rows']}`",
+        f"- 做多收益来源: `{result['edge_source_by_side'].get('LONG')}`",
+        f"- 做空收益来源: `{result['edge_source_by_side'].get('SHORT')}`",
         "",
     ]
     if candidates.empty: