Add Entry opportunity training diagnostics

training/trader_training/training.py

@@ -89,7 +89,7 @@ def train_small_models(args: Any) -> None:
     model_manifest: dict[str, Any] = {}
     for model_name, spec in TARGETS.items():
         dataset = read_parquet(root / "dataset" / spec["dataset"])
-        if model_name == "ENTRY" and _conditional_entry_enabled(args):
+        if model_name == "ENTRY" and _conditional_entry_source(args) == "direction_label":
             dataset = _attach_direction_fit_labels(root, dataset)
         if args.max_rows and len(dataset) > args.max_rows:
             dataset = dataset.sort_values("event_time").tail(args.max_rows).copy()
@@ -189,7 +189,17 @@ def train_small_models(args: Any) -> None:
 
 
 def _conditional_entry_enabled(args: Any) -> bool:
-    return bool(getattr(args, "conditional_entry_direction_labels", False))
+    return _conditional_entry_source(args) != "none"
+
+
+def _conditional_entry_source(args: Any) -> str:
+    source = str(getattr(args, "conditional_entry_source", "none") or "none").strip().lower()
+    if bool(getattr(args, "conditional_entry_direction_labels", False)):
+        source = "direction_label"
+    allowed = {"none", "direction_label", "side_opportunity"}
+    if source not in allowed:
+        raise ValueError(f"unsupported conditional Entry source: {source}")
+    return source
 
 
 def _attach_direction_fit_labels(root: Path, entry_dataset: pd.DataFrame) -> pd.DataFrame:
@@ -213,19 +223,29 @@ def _attach_direction_fit_labels(root: Path, entry_dataset: pd.DataFrame) -> pd.
 
 
 def _head_train_mask(model_name: str, head_name: str, train: pd.DataFrame, args: Any) -> tuple[np.ndarray, str]:
-    if model_name != "ENTRY" or not _conditional_entry_enabled(args):
+    source = _conditional_entry_source(args)
+    if model_name != "ENTRY" or source == "none":
         return np.ones(len(train), dtype=bool), "ALL_FIT_ROWS"
     if head_name.startswith("long_"):
-        condition_column = "long_target"
-        filter_name = "DIRECTION_LABEL_LONG_FIT_ROWS"
+        side = "LONG"
+        direction_label_column = "long_target"
+        opportunity_column = "long_max_achievable_net_edge_bps"
     elif head_name.startswith("short_"):
-        condition_column = "short_target"
-        filter_name = "DIRECTION_LABEL_SHORT_FIT_ROWS"
+        side = "SHORT"
+        direction_label_column = "short_target"
+        opportunity_column = "short_max_achievable_net_edge_bps"
     else:
         return np.ones(len(train), dtype=bool), "ALL_FIT_ROWS"
-    if condition_column not in train.columns:
-        raise ValueError(f"conditional Entry training requires {condition_column} for head {head_name}")
-    mask = pd.to_numeric(train[condition_column], errors="coerce").fillna(0).astype(int).eq(1).to_numpy()
+    if source == "direction_label":
+        if direction_label_column not in train.columns:
+            raise ValueError(f"conditional Entry training requires {direction_label_column} for head {head_name}")
+        mask = pd.to_numeric(train[direction_label_column], errors="coerce").fillna(0).astype(int).eq(1).to_numpy()
+        return mask, f"DIRECTION_LABEL_{side}_FIT_ROWS"
+    threshold = float(getattr(args, "conditional_entry_opportunity_bps", 40.0) or 40.0)
+    if opportunity_column not in train.columns:
+        raise ValueError(f"side opportunity Entry training requires {opportunity_column} for head {head_name}")
+    mask = pd.to_numeric(train[opportunity_column], errors="coerce").ge(threshold).fillna(False).to_numpy()
+    filter_name = f"SIDE_OPPORTUNITY_{side}_GE_{threshold:g}_BPS_FIT_ROWS"
     return mask, filter_name