fix: train valid purged ML artifacts

ml/optimizer.py

@@ -156,6 +156,14 @@ def purged_time_series_splits(rows, n_splits=VALIDATION_SPLITS,
             yield np.array(purged_train, dtype=int), np.array(val_idx, dtype=int)
 
 
+def viable_classification_splits(y, splits):
+    """Yield only folds whose training window contains both target classes."""
+    for train_idx, val_idx in splits:
+        if len(np.unique(y[train_idx])) < 2:
+            continue
+        yield train_idx, val_idx
+
+
 def _build_model():
     return GradientBoostingClassifier(
         n_estimators=300,
@@ -377,12 +385,12 @@ def train_model(rows):
     cv_recall = []
     fold_results = []
 
-    splits = list(purged_time_series_splits(
+    splits = list(viable_classification_splits(y, purged_time_series_splits(
         labeled,
         n_splits=VALIDATION_SPLITS,
         label_horizon_days=LABEL_HORIZON_DAYS,
         embargo_days=0,
-    ))
+    )))
     if not splits:
         log.error("No viable purged validation splits. Need more history for %dd label horizon.",
                   LABEL_HORIZON_DAYS)

tests/test_ml_optimizer_validation.py

@@ -1,5 +1,7 @@
 from datetime import datetime, timedelta
 
+import numpy as np
+
 from ml import optimizer
 
 
@@ -72,3 +74,16 @@ def test_run_out_of_sample_comparison_scores_only_validation_rows_with_fold_weig
 
     caution_equal = next(bucket for bucket in comparison["equal_weight"] if bucket["label"] == "CAUTION — OVERHEATED")
     assert caution_equal["days"] == 2
+
+
+def test_classification_splits_skip_training_windows_with_one_class():
+    y = np.array([1, 1, 1, 0, 1, 0])
+    splits = [
+        (np.array([0, 1]), np.array([2, 3])),
+        (np.array([0, 1, 3, 4]), np.array([5])),
+    ]
+
+    viable = list(optimizer.viable_classification_splits(y, splits))
+
+    assert len(viable) == 1
+    assert viable[0][0].tolist() == [0, 1, 3, 4]