Improve Trader entry quality training diagnostics

training/trader_training/price_plan_search.py

@@ -194,7 +194,7 @@ def _plan_side_rows(
     gross = np.where(target_first, target_bps, np.where(stop_first, -stop_bps, timeout_return))
     price_plan_net = gross - cost_bps
     expected_net = max_achievable_gross - cost_bps
-    positive = expected_net >= min_expected_edge_bps
+    positive = price_plan_net >= min_expected_edge_bps
     ambiguous = target_any & stop_any & (target_index == stop_index)
 
     rows: list[dict[str, Any]] = []
@@ -264,21 +264,25 @@ def _plan_summary(rows: pd.DataFrame) -> pd.DataFrame:
     split_rows["avg_price_plan_edge_eval"] = split_rows[
         [f"avg_price_plan_net_edge_bps_{split}" for split in (TUNE_SPLIT, VALIDATION_LOCKED_SPLIT, LATEST_STRESS_SPLIT)]
     ].mean(axis=1)
+    split_rows["min_price_plan_edge_eval"] = split_rows[
+        [f"avg_price_plan_net_edge_bps_{split}" for split in (TUNE_SPLIT, VALIDATION_LOCKED_SPLIT, LATEST_STRESS_SPLIT)]
+    ].min(axis=1)
     split_rows["min_margin_eval"] = split_rows[
         [f"target_rate_margin_{split}" for split in (TUNE_SPLIT, VALIDATION_LOCKED_SPLIT, LATEST_STRESS_SPLIT)]
     ].min(axis=1)
-    # The search score is not an上线门槛. It only chooses the next experiment:
-    # enough positive samples, less negative average edge, and stable behavior
-    # across tune/validation/stress.
+    # The search score is not an上线门槛. It only chooses the next experiment.
+    # Fixed price-plan net edge is weighted most heavily; max future edge is
+    # retained only as a weak tie-breaker because it can be optimistic.
     positive_rate_penalty = (
-        (0.08 - split_rows["min_positive_label_rate_eval"]).clip(lower=0.0) * 80.0
-        + (split_rows["max_positive_label_rate_eval"] - 0.45).clip(lower=0.0) * 30.0
+        (0.03 - split_rows["min_positive_label_rate_eval"]).clip(lower=0.0) * 120.0
+        + (split_rows["max_positive_label_rate_eval"] - 0.55).clip(lower=0.0) * 20.0
     )
     spread_bonus = np.log1p((split_rows["target_bps"] - split_rows["stop_bps"]).clip(lower=0.0))
     split_rows["score"] = (
-        split_rows["avg_edge_eval"]
-        + split_rows["avg_price_plan_edge_eval"] * 0.5
-        + split_rows["min_margin_eval"] * 20.0
+        split_rows["avg_price_plan_edge_eval"] * 8.0
+        + split_rows["min_price_plan_edge_eval"] * 3.0
+        + split_rows["min_margin_eval"] * 80.0
+        + split_rows["avg_edge_eval"] * 0.05
         - positive_rate_penalty
         + spread_bonus
     )
@@ -287,8 +291,10 @@ def _plan_summary(rows: pd.DataFrame) -> pd.DataFrame:
 
 def _select_best_plan(summary: pd.DataFrame) -> dict[str, Any]:
     candidates = summary[
-        (summary["min_positive_label_rate_eval"] >= 0.08)
-        & (summary["max_positive_label_rate_eval"] <= 0.45)
+        (summary["min_positive_label_rate_eval"] >= 0.03)
+        & (summary["max_positive_label_rate_eval"] <= 0.55)
+        & (summary["avg_price_plan_edge_eval"] > 0.0)
+        & (summary["min_price_plan_edge_eval"] > -1.0)
         & (summary["target_bps"] > summary["stop_bps"])
     ]
     if candidates.empty:
@@ -305,6 +311,7 @@ def _select_best_plan(summary: pd.DataFrame) -> dict[str, Any]:
         "score": float(row["score"]),
         "avg_edge_eval": float(row["avg_edge_eval"]),
         "avg_price_plan_edge_eval": float(row["avg_price_plan_edge_eval"]),
+        "min_price_plan_edge_eval": float(row["min_price_plan_edge_eval"]),
         "min_margin_eval": float(row["min_margin_eval"]),
         "min_positive_label_rate_eval": float(row["min_positive_label_rate_eval"]),
         "max_positive_label_rate_eval": float(row["max_positive_label_rate_eval"]),
@@ -332,7 +339,7 @@ def _markdown_report(payload: dict[str, Any], summary: pd.DataFrame) -> str:
         "",
         _markdown_table(top),
         "",
-        "说明：positive_label_rate 和 avg_expected_net_edge_bps 按“未来窗口最大可拿净收益”统计；target_hit_rate、stop_hit_rate、avg_price_plan_net_edge_bps 只用来检查固定止盈止损计划是否顺手。这里选的是下一轮实验用的价格计划，不是上线结论。真正能不能上线仍然看模型训练、PM 搜索、validation_locked 和 latest_stress 回测。",
+        "说明：positive_label_rate 和 avg_price_plan_net_edge_bps 按固定止盈止损计划统计；avg_expected_net_edge_bps 只是辅助观察未来最大可拿空间，不能单独决定价格计划。这里选的是下一轮实验用的价格计划，不是上线结论。真正能不能上线仍然看模型训练、PM 搜索、validation_locked 和 latest_stress 回测。",
         "",
     ]
     return "\n".join(lines)