spforge 0.8.8__py3-none-any.whl → 0.8.18__py3-none-any.whl

spforge/ratings/utils.py

@@ -2,6 +2,10 @@ import polars as pl
 
 from spforge.data_structures import ColumnNames
 
+# Internal column names for scaled participation weights
+_SCALED_PW = "__scaled_participation_weight__"
+_SCALED_PPW = "__scaled_projected_participation_weight__"
+
 
 def add_team_rating(
     df: pl.DataFrame,
@@ -46,11 +50,14 @@ def add_team_rating_projected(
     tid = column_names.team_id
     ppw = column_names.projected_participation_weight
 
-    if ppw:
+    # Use scaled column if available (clipped to [0, 1]), otherwise raw column
+    weight_col = _SCALED_PPW if _SCALED_PPW in df.columns else ppw
+
+    if weight_col and weight_col in df.columns:
         return df.with_columns(
             (
-                (pl.col(ppw) * pl.col(player_rating_col)).sum().over([mid, tid])
-                / pl.col(ppw).sum().over([mid, tid])
+                (pl.col(weight_col) * pl.col(player_rating_col)).sum().over([mid, tid])
+                / pl.col(weight_col).sum().over([mid, tid])
             ).alias(team_rating_out)
         )
 
@@ -118,11 +125,14 @@ def add_rating_mean_projected(
     mid = column_names.match_id
     ppw = column_names.projected_participation_weight
 
-    if ppw:
+    # Use scaled column if available (clipped to [0, 1]), otherwise raw column
+    weight_col = _SCALED_PPW if _SCALED_PPW in df.columns else ppw
+
+    if weight_col and weight_col in df.columns:
         return df.with_columns(
             (
-                (pl.col(ppw) * pl.col(player_rating_col)).sum().over(mid)
-                / pl.col(ppw).sum().over(mid)
+                (pl.col(weight_col) * pl.col(player_rating_col)).sum().over(mid)
+                / pl.col(weight_col).sum().over(mid)
             ).alias(rating_mean_out)
         )
 

spforge/scorer/_score.py

@@ -366,18 +366,49 @@ class PWMSE(BaseScorer):
         self.labels = labels
         self.evaluation_labels = evaluation_labels
 
+        self._needs_extension = False
+        self._needs_slicing = False
         self._eval_indices: list[int] | None = None
+        self._extension_mapping: dict[int, int] | None = None
+
         if self.evaluation_labels is not None and self.labels is not None:
-            label_to_idx = {lbl: i for i, lbl in enumerate(self.labels)}
-            self._eval_indices = [label_to_idx[lbl] for lbl in self.evaluation_labels]
+            training_set = set(self.labels)
+            eval_set = set(self.evaluation_labels)
+
+            if eval_set <= training_set:
+                self._needs_slicing = True
+                label_to_idx = {lbl: i for i, lbl in enumerate(self.labels)}
+                self._eval_indices = [label_to_idx[lbl] for lbl in self.evaluation_labels]
+            elif training_set <= eval_set:
+                self._needs_extension = True
+                eval_label_to_idx = {lbl: i for i, lbl in enumerate(self.evaluation_labels)}
+                self._extension_mapping = {
+                    train_idx: eval_label_to_idx[lbl]
+                    for train_idx, lbl in enumerate(self.labels)
+                }
+            else:
+                raise ValueError(
+                    f"evaluation_labels must be a subset or superset of labels. "
+                    f"labels={self.labels}, evaluation_labels={self.evaluation_labels}"
+                )
+
+    def _align_predictions(self, preds: np.ndarray) -> np.ndarray:
+        if self._needs_slicing and self._eval_indices is not None:
+            sliced = preds[:, self._eval_indices]
+            row_sums = sliced.sum(axis=1, keepdims=True)
+            row_sums = np.where(row_sums == 0, 1.0, row_sums)
+            return sliced / row_sums
+
+        if self._needs_extension and self._extension_mapping is not None:
+            n_samples = preds.shape[0]
+            n_eval_labels = len(self.evaluation_labels)
+            extended = np.full((n_samples, n_eval_labels), 1e-5, dtype=np.float64)
+            for train_idx, eval_idx in self._extension_mapping.items():
+                extended[:, eval_idx] = preds[:, train_idx]
+            row_sums = extended.sum(axis=1, keepdims=True)
+            return extended / row_sums
 
-    def _slice_and_renormalize(self, preds: np.ndarray) -> np.ndarray:
-        if self._eval_indices is None:
-            return preds
-        sliced = preds[:, self._eval_indices]
-        row_sums = sliced.sum(axis=1, keepdims=True)
-        row_sums = np.where(row_sums == 0, 1.0, row_sums)
-        return sliced / row_sums
+        return preds
 
     def _get_scoring_labels(self) -> list[int]:
         if self.evaluation_labels is not None:
@@ -446,7 +477,7 @@ class PWMSE(BaseScorer):
 
                 targets = gran_df[self.target].to_numpy().astype(np.float64)
                 preds = np.asarray(gran_df[self.pred_column].to_list(), dtype=np.float64)
-                preds = self._slice_and_renormalize(preds)
+                preds = self._align_predictions(preds)
                 score = self._pwmse_score(targets, preds)
                 if self.compare_to_naive:
                     naive_probs_list = _naive_probability_predictions_for_df(
@@ -464,7 +495,7 @@ class PWMSE(BaseScorer):
 
         targets = df[self.target].to_numpy().astype(np.float64)
         preds = np.asarray(df[self.pred_column].to_list(), dtype=np.float64)
-        preds = self._slice_and_renormalize(preds)
+        preds = self._align_predictions(preds)
         score = self._pwmse_score(targets, preds)
         if self.compare_to_naive:
             naive_probs_list = _naive_probability_predictions_for_df(

spforge/transformers/_other_transformer.py

@@ -8,8 +8,9 @@ from sklearn.base import BaseEstimator, TransformerMixin
 
 
 class GroupByReducer(BaseEstimator, TransformerMixin):
-    def __init__(self, granularity: list[str]):
+    def __init__(self, granularity: list[str], aggregation_weight: str | None = None):
         self.granularity = granularity
+        self.aggregation_weight = aggregation_weight
 
     @nw.narwhalify
     def fit(self, X: IntoFrameT, y: Any = None):
@@ -26,18 +27,47 @@ class GroupByReducer(BaseEstimator, TransformerMixin):
             raise ValueError("Could not find granularity columns in dataframe %s", self.granularity)
 
         non_keys = [c for c in df.columns if c not in keys]
-        num_cols = [c for c in non_keys if pd.api.types.is_numeric_dtype(df[c])]
+        schema = df.schema
+        num_cols = [c for c in non_keys if schema[c].is_numeric()]
         other_cols = [c for c in non_keys if c not in num_cols]
 
         aggs: list[nw.Expr] = []
 
+        # Backwards compatibility: old pickled objects may not have aggregation_weight
+        weight_col = getattr(self, "aggregation_weight", None)
+        has_weight = weight_col and weight_col in df.columns
+
         for c in num_cols:
-            aggs.append(nw.col(c).mean().alias(c))
+            if c == weight_col:
+                aggs.append(nw.col(c).sum().alias(c))
+            elif has_weight:
+                aggs.append((nw.col(c) * nw.col(weight_col)).sum().alias(f"__{c}_weighted_sum"))
+                aggs.append(nw.col(c).mean().alias(f"__{c}_fallback"))
+            else:
+                aggs.append(nw.col(c).mean().alias(c))
 
         for c in other_cols:
             aggs.append(nw.col(c).first().alias(c))
 
+        if has_weight:
+            aggs.append(nw.col(weight_col).sum().alias("__weight_sum"))
+
         out = df.group_by(keys).agg(aggs)
+
+        if has_weight:
+            weighted_cols = [c for c in num_cols if c != weight_col]
+            for c in weighted_cols:
+                out = out.with_columns(
+                    nw.when((~nw.col("__weight_sum").is_null()) & (nw.col("__weight_sum") != 0))
+                    .then(nw.col(f"__{c}_weighted_sum") / nw.col("__weight_sum"))
+                    .otherwise(nw.col(f"__{c}_fallback"))
+                    .alias(c)
+                )
+            drop_cols = [f"__{c}_weighted_sum" for c in weighted_cols]
+            drop_cols += [f"__{c}_fallback" for c in weighted_cols]
+            drop_cols.append("__weight_sum")
+            out = out.drop(drop_cols)
+
         return out
 
     @nw.narwhalify
@@ -59,12 +89,12 @@ class GroupByReducer(BaseEstimator, TransformerMixin):
         if sample_weight is not None:
             df = df.with_columns(nw.lit(sample_weight).alias("__sw"))
 
-        y_is_numeric = df.select(nw.col("__y")).schema["__y"].is_numeric()
+        y_uniques = df.group_by(keys).agg(nw.col("__y").n_unique().alias("__y_nunique"))
+        non_uniform = y_uniques.filter(nw.col("__y_nunique") > 1)
+        if len(non_uniform) > 0:
+            raise ValueError("Target (y) must be uniform within each granularity group")
 
-        if y_is_numeric:
-            agg_exprs = [nw.col("__y").mean().alias("__y")]
-        else:
-            agg_exprs = [nw.col("__y").first().alias("__y")]
+        agg_exprs = [nw.col("__y").first().alias("__y")]
 
         if sample_weight is not None:
             agg_exprs.append(nw.col("__sw").sum().alias("__sw"))

{spforge-0.8.8.dist-info → spforge-0.8.18.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: spforge
-Version: 0.8.8
+Version: 0.8.18
 Summary: A flexible framework for generating features, ratings, and building machine learning or other models for training and inference on sports data.
 Author-email: Mathias Holmstrøm <mathiasholmstom@gmail.com>
 License: See LICENSE file

{spforge-0.8.8.dist-info → spforge-0.8.18.dist-info}/RECORD

@@ -14,7 +14,7 @@ examples/nba/data/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU
 examples/nba/data/game_player_subsample.parquet,sha256=ODJxHC-mUYbJ7r-ScUFtPU7hrFuxLUbbDSobmpCkw0w,279161
 examples/nba/data/utils.py,sha256=41hxLQ1d6ZgBEcHa5MI0-fG5KbsRi07cclMPQZM95ek,509
 spforge/__init__.py,sha256=8vZhy7XUpzqWkVKpXqwqOLDkQlNytRhyf4qjwObfXgU,468
-spforge/autopipeline.py,sha256=ZUwv6Q6O8cD0u5TiSqG6lhW0j16RlSb160AzuOeL2R8,23186
+spforge/autopipeline.py,sha256=rZ6FhJxcgNLvtr3hTVkEiW4BiorgXxADThfMuQ42orE,29866
 spforge/base_feature_generator.py,sha256=RbD00N6oLCQQcEb_VF5wbwZztl-X8k9B0Wlaj9Os1iU,668
 spforge/data_structures.py,sha256=k82v5r79vl0_FAVvsxVF9Nbzb5FoHqVrlHZlEXGc5gQ,7298
 spforge/features_generator_pipeline.py,sha256=n8vzZKqXNFcFRDWZhllnkhAh5NFXdOD3FEIOpHcay8E,8208
@@ -30,7 +30,7 @@ spforge/estimator/__init__.py,sha256=zIJ4u7WGPOALPx8kVBppBOqklI4lQPl9QBWT8JjjFoY
 spforge/estimator/_conditional_estimator.py,sha256=JSHpOg5lv3kRv_VzSZ0fKbwCO2dJv9XpyLs9lS81psU,4904
 spforge/estimator/_frequency_bucketing_classifier.py,sha256=d7wDpOCoKWf-WoXtzwahjtmAozkFdKE3-pzs477WMYc,6055
 spforge/estimator/_granularity_estimator.py,sha256=pUNmtpDFoOVbS9mHfO-zvidPIKJgWts0y2VnhJ8VWww,3829
-spforge/estimator/_group_by_estimator.py,sha256=aXuDvRWvvgK4SEI_DMYscvathmPb6nkMxnqKgG8HC0Y,2769
+spforge/estimator/_group_by_estimator.py,sha256=o-xv_PJJyWBaKv5Eo4EPbOvb9i0CuebZnX4GtEFp_Js,3120
 spforge/estimator/_ordinal_classifier.py,sha256=j_dfVHeX-6eZgPwwsYbkbP6bPrKH2a5S-N8vfP5hneA,1993
 spforge/estimator/_sklearn_enhancer_estimator.py,sha256=DZ-UlmeazXPd6uEnlbVv79syZ5FPa64voUyKArtjjUs,4664
 spforge/feature_generator/__init__.py,sha256=wfLfUkC_lLOCpy7NgDytK-l3HUAuhikuQXdKCgSGbuA,556
@@ -47,32 +47,32 @@ spforge/hyperparameter_tuning/__init__.py,sha256=N2sKG4SvG41hlsFT2kx_DQYMmXsQr-8
 spforge/hyperparameter_tuning/_default_search_spaces.py,sha256=Sm5IrHAW0-vRC8jqCPX0pDi_C-W3L_MoEKGA8bx1Zbc,7546
 spforge/hyperparameter_tuning/_tuner.py,sha256=uovhGqhe8-fdhi79aErUmE2h5NCycFQEIRv5WCjpC7E,16732
 spforge/performance_transformers/__init__.py,sha256=U6d7_kltbUMLYCGBk4QAFVPJTxXD3etD9qUftV-O3q4,422
-spforge/performance_transformers/_performance_manager.py,sha256=KwAga6dGhNkXi-MDW6LPjwk6VZwCcjo5L--jnk9aio8,9706
+spforge/performance_transformers/_performance_manager.py,sha256=WmjmlMEnq7y75MiI_s9Y-9eMXIyhPTUKrwsXRtgYp0k,9620
 spforge/performance_transformers/_performances_transformers.py,sha256=0lxuWjAfWBRXRgQsNJHjw3P-nlTtHBu4_bOVdoy7hq4,15536
 spforge/ratings/__init__.py,sha256=OZVH2Lo6END3n1X8qi4QcyAPlThIwAYwVKCiIuOQSQU,576
 spforge/ratings/_base.py,sha256=dRMkIGj5-2zKddygaEA4g16WCyXon7v8Xa1ymm7IuoM,14335
-spforge/ratings/_player_rating.py,sha256=MyqsyLSY6d7_bxDSnF8eWOyXpSCADWGdepdFSGM4cHw,51365
+spforge/ratings/_player_rating.py,sha256=JSTXdaRw_b8ZoZxgmMnZrYG7gPg8GKawqalLd16SK1M,56066
 spforge/ratings/_team_rating.py,sha256=T0kFiv3ykYSrVGGsVRa8ZxLB0WMnagxqdFDzl9yZ_9g,24813
 spforge/ratings/enums.py,sha256=s7z_RcZS6Nlgfa_6tasO8_IABZJwywexe7sep9DJBgo,1739
 spforge/ratings/league_identifier.py,sha256=_KDUKOwoNU6RNFKE5jju4eYFGVNGBdJsv5mhNvMakfc,6019
 spforge/ratings/league_start_rating_optimizer.py,sha256=Q4Vo3QT-r55qP4aD9WftsTB00UOSRvxM1khlyuAGWNM,8582
 spforge/ratings/player_performance_predictor.py,sha256=cMxzQuk0nF1MsT_M32g-3mxVdAEbZ-S7TUjEPYdo3Yg,8361
-spforge/ratings/start_rating_generator.py,sha256=_7hIJ9KRVCwsCoY1GIzY8cuOdHR8RH_BCMeMwQG3E04,6776
+spforge/ratings/start_rating_generator.py,sha256=eSasa5Oe9n4IoTGjFCYyFQAGrJtzrBW-Qor97lmaYuM,6776
 spforge/ratings/team_performance_predictor.py,sha256=ThQOmYQUqKBB46ONYHOMM2arXFH8AkyKpAZzs80SjHA,7217
-spforge/ratings/team_start_rating_generator.py,sha256=ZJe84sTvE4Yep3d4wKJMMJn2Q4PhcCwkO7Wyd5nsYUA,5110
-spforge/ratings/utils.py,sha256=qms5J5SD-FyXDR2G8giDMbu_AoLgI135pjW4nghxROg,3940
+spforge/ratings/team_start_rating_generator.py,sha256=vK-_m8KwcHopchch_lKNHSGLiiNm5q9Lenm0d1cP_po,5110
+spforge/ratings/utils.py,sha256=_zFemqz2jJkH8rn2EZpDt8N6FELUmYp9qCnPzRtOIGU,4497
 spforge/scorer/__init__.py,sha256=wj8PCvYIl6742Xwmt86c3oy6iqE8Ss-OpwHud6kd9IY,256
-spforge/scorer/_score.py,sha256=TR0T9nJj0aeVgGfOE0fZmXlO66CELulYwxhi7ZAxhvY,56184
+spforge/scorer/_score.py,sha256=kNuqiK3F5mUEAVD7KjWYY7E_AkRrspR362QBm_jyElg,57623
 spforge/transformers/__init__.py,sha256=IPCsMcsgBqG52d0ttATLCY4HvFCQZddExlLt74U-zuI,390
 spforge/transformers/_base.py,sha256=-smr_McQF9bYxM5-Agx6h7Xv_fhZzPfpAdQV-qK18bs,1134
 spforge/transformers/_net_over_predicted.py,sha256=5dC8pvA1DNO0yXPSgJSMGU8zAHi-maUELm7FqFQVo-U,2321
 spforge/transformers/_operator.py,sha256=jOH7wdMBLg6R2hlH_FU6eA0gjs-Q0vFimTo7fXgKpjI,2964
-spforge/transformers/_other_transformer.py,sha256=xLfaFIhkFsigAoitB4x3F8An2j9ymdjQy5VrsTvJlrA,3152
+spforge/transformers/_other_transformer.py,sha256=w2a7Wnki3vJe4GAkSa4kealw0GILIo6nE_9-3M10owA,4646
 spforge/transformers/_predictor.py,sha256=2sE6gfVrilXzPVcBurSrtqHw33v2ljygQcEYXt9LhZc,3119
 spforge/transformers/_simple_transformer.py,sha256=zGUFNQYMeoDSa2CoQejQNiNmKCBN5amWTvyOchiUHj0,5660
 spforge/transformers/_team_ratio_predictor.py,sha256=g8_bR53Yyv0iNCtol1O9bgJSeZcIco_AfbQuUxQJkeY,6884
-spforge-0.8.8.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-tests/test_autopipeline.py,sha256=WXHeqBdjQD6xaXVkzvS8ocz0WVP9R7lN0PiHJ2iD8nA,16911
+spforge-0.8.18.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+tests/test_autopipeline.py,sha256=7cNAn-nmGolfyfk3THh9IKcHZfRA-pLYC_xAyMg-No4,26863
 tests/test_autopipeline_context.py,sha256=IuRUY4IA6uMObvbl2pXSaXO2_tl3qX6wEbTZY0dkTMI,1240
 tests/test_feature_generator_pipeline.py,sha256=CK0zVL8PfTncy3RmG9i-YpgwjOIV7yJhV7Q44tbetI8,19020
 tests/cross_validator/test_cross_validator.py,sha256=itCGhNY8-NbDbKbhxHW20wiLuRst7-Rixpmi3FSKQtA,17474
@@ -92,12 +92,14 @@ tests/feature_generator/test_rolling_mean_days.py,sha256=EyOvdJDnmgPfe13uQBOkwo7
 tests/feature_generator/test_rolling_window.py,sha256=YBJo36OK3ILYeXrH06ylXqviUcCaGYaVQaK5RJzwM7Y,23239
 tests/hyperparameter_tuning/test_estimator_tuner.py,sha256=iewME41d6LR2aQ0OtohGFtN_ocJUwTeqvs6L0QDmfG4,4413
 tests/hyperparameter_tuning/test_rating_tuner.py,sha256=PyCFP3KPc4Iy9E_X9stCVxra14uMgC1tuRwuQ30rO_o,13195
-tests/performance_transformers/test_performance_manager.py,sha256=bfC5GiBuzHw-mLmKeEzBUUPuKm0ayax2bsF1j88W8L0,10120
+tests/performance_transformers/test_performance_manager.py,sha256=gjuuV_hb27kCo_kUecPKG3Cbot2Gqis1W3kw2A4ovS4,10690
 tests/performance_transformers/test_performances_transformers.py,sha256=A-tGiCx7kXrj1cVj03Bc7prOeZ1_Ryz8YFx9uj3eK6w,11064
-tests/ratings/test_player_rating_generator.py,sha256=FGH3Tq0uFoSlkS_XMldsUKhsovBRBvzH9EbqjKvg2O0,59601
+tests/ratings/test_player_rating_generator.py,sha256=SKLaBQBsHYslc2Nia2AxZ8A9Cy16MbZAWjLyOjvcMnA,64094
+tests/ratings/test_player_rating_no_mutation.py,sha256=GzO3Hl__5K68DS3uRLefwnbcTJOvBM7cZqww4M21UZM,8493
 tests/ratings/test_ratings_property.py,sha256=ckyfGILXa4tfQvsgyXEzBDNr2DUmHwFRV13N60w66iE,6561
 tests/ratings/test_team_rating_generator.py,sha256=cDnf1zHiYC7pkgydE3MYr8wSTJIq-bPfSqhIRI_4Tic,95357
-tests/scorer/test_score.py,sha256=_Vd6tKpy_1GeOxU7Omxci4CFf7PvRGMefEI0gv2gV6A,74688
+tests/ratings/test_utils_scaled_weights.py,sha256=iHxe6ZDUB_I2B6HT0xTGqXBkl7gRlqVV0e_7Lwun5po,4988
+tests/scorer/test_score.py,sha256=rw3xJs6xqWVpalVMUQz557m2JYGR7PmhrsjfTex0b0c,79121
 tests/scorer/test_score_aggregation_granularity.py,sha256=h-hyFOLzwp-92hYVU7CwvlRJ8jhB4DzXCtqgI-zcoqM,13677
 tests/transformers/test_estimator_transformer_context.py,sha256=5GOHbuWCWBMFwwOTJOuD4oNDsv-qDR0OxNZYGGuMdag,1819
 tests/transformers/test_net_over_predicted.py,sha256=vh7O1iRRPf4vcW9aLhOMAOyatfM5ZnLsQBKNAYsR3SU,3363
@@ -105,7 +107,7 @@ tests/transformers/test_other_transformer.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRk
 tests/transformers/test_predictor_transformer.py,sha256=N1aBYLjN3ldpYZLwjih_gTFYSMitrZu-PNK78W6RHaQ,6877
 tests/transformers/test_simple_transformer.py,sha256=wWR0qjLb_uS4HXrJgGdiqugOY1X7kwd1_OPS02IT2b8,4676
 tests/transformers/test_team_ratio_predictor.py,sha256=fOUP_JvNJi-3kom3ZOs1EdG0I6Z8hpLpYKNHu1eWtOw,8562
-spforge-0.8.8.dist-info/METADATA,sha256=fO2JHqnnqOrjkWZ1Zh4rgYg58bi4YzxhSa8I72wqDs4,20047
-spforge-0.8.8.dist-info/WHEEL,sha256=qELbo2s1Yzl39ZmrAibXA2jjPLUYfnVhUNTlyF1rq0Y,92
-spforge-0.8.8.dist-info/top_level.txt,sha256=6UW2M5a7WKOeaAi900qQmRKNj5-HZzE8-eUD9Y9LTq0,23
-spforge-0.8.8.dist-info/RECORD,,
+spforge-0.8.18.dist-info/METADATA,sha256=54l0UTrew2ot0_4k22hLKL-oXbQ4hlA1_KAXIqf_umw,20048
+spforge-0.8.18.dist-info/WHEEL,sha256=wUyA8OaulRlbfwMtmQsvNngGrxQHAvkKcvRmdizlJi0,92
+spforge-0.8.18.dist-info/top_level.txt,sha256=6UW2M5a7WKOeaAi900qQmRKNj5-HZzE8-eUD9Y9LTq0,23
+spforge-0.8.18.dist-info/RECORD,,

{spforge-0.8.8.dist-info → spforge-0.8.18.dist-info}/WHEEL

@@ -1,5 +1,5 @@
 Wheel-Version: 1.0
-Generator: setuptools (80.10.1)
+Generator: setuptools (80.10.2)
 Root-Is-Purelib: true
 Tag: py3-none-any
 

tests/performance_transformers/test_performance_manager.py

@@ -56,6 +56,21 @@ def test_performance_weights_manager_basic_flow(sample_data):
     assert output_df["weighted_performance"].iloc[0] == pytest.approx(0.6)
 
 
+def test_performance_weights_manager_keeps_mean_when_weights_not_normalized():
+    df = pd.DataFrame(
+        {
+            "feat_a": [0.0, 1.0, 2.0, 3.0],
+            "feat_b": [3.0, 2.0, 1.0, 0.0],
+        }
+    )
+    weights = [ColumnWeight(name="feat_a", weight=0.9), ColumnWeight(name="feat_b", weight=0.5)]
+
+    manager = PerformanceWeightsManager(weights=weights, transformer_names=["min_max"], prefix="")
+    output_df = nw.from_native(manager.fit_transform(df)).to_pandas()
+
+    assert output_df["weighted_performance"].mean() == pytest.approx(0.5, abs=1e-6)
+
+
 def test_lower_is_better_logic():
     df = pd.DataFrame({"feat_a": [1.0, 0.0]})
     weights = [ColumnWeight(name="feat_a", weight=1.0, lower_is_better=True)]

tests/ratings/test_player_rating_generator.py

@@ -551,6 +551,63 @@ def test_fit_transform_scales_participation_weight_by_fit_quantile(base_cn):
     assert p1_change / p2_change == pytest.approx(expected_ratio, rel=1e-6)
 
 
+def test_fit_transform_auto_scales_participation_weight_when_out_of_bounds(base_cn):
+    """Automatically enable scaling when participation weights exceed [0, 1]."""
+    df = pl.DataFrame(
+        {
+            "pid": ["P1", "P2", "O1", "O2"],
+            "tid": ["T1", "T1", "T2", "T2"],
+            "mid": ["M1", "M1", "M1", "M1"],
+            "dt": ["2024-01-01"] * 4,
+            "perf": [0.9, 0.9, 0.1, 0.1],
+            "pw": [10.0, 20.0, 10.0, 10.0],
+        }
+    )
+    gen = PlayerRatingGenerator(
+        performance_column="perf",
+        column_names=base_cn,
+        auto_scale_performance=True,
+        start_harcoded_start_rating=1000.0,
+    )
+    gen.fit_transform(df)
+
+    start_rating = 1000.0
+    p1_change = gen._player_off_ratings["P1"].rating_value - start_rating
+    p2_change = gen._player_off_ratings["P2"].rating_value - start_rating
+
+    q = df["pw"].quantile(0.99, "linear")
+    expected_ratio = min(1.0, 10.0 / q) / min(1.0, 20.0 / q)
+
+    assert gen.scale_participation_weights is True
+    assert p1_change / p2_change == pytest.approx(expected_ratio, rel=1e-6)
+
+
+def test_fit_transform_auto_scale_logs_warning_when_out_of_bounds(base_cn, caplog):
+    """Auto-scaling should emit a warning when participation weights exceed [0, 1]."""
+    df = pl.DataFrame(
+        {
+            "pid": ["P1", "P2", "O1", "O2"],
+            "tid": ["T1", "T1", "T2", "T2"],
+            "mid": ["M1", "M1", "M1", "M1"],
+            "dt": ["2024-01-01"] * 4,
+            "perf": [0.9, 0.9, 0.1, 0.1],
+            "pw": [10.0, 20.0, 10.0, 10.0],
+        }
+    )
+    gen = PlayerRatingGenerator(
+        performance_column="perf",
+        column_names=base_cn,
+        auto_scale_performance=True,
+        start_harcoded_start_rating=1000.0,
+    )
+    with caplog.at_level("WARNING"):
+        gen.fit_transform(df)
+
+    assert any(
+        "Auto-scaling participation weights" in record.message for record in caplog.records
+    )
+
+
 def test_future_transform_scales_projected_participation_weight_by_fit_quantile():
     """Future projected participation weights should scale with fit quantile and be clipped."""
     cn = ColumnNames(
@@ -1689,3 +1746,73 @@ def test_fit_transform__player_rating_difference_from_team_projected_feature(bas
     for row in result.iter_rows(named=True):
         expected = row[player_col] - row[team_col]
         assert row[diff_col] == pytest.approx(expected, rel=1e-9)
+
+
+def test_fit_transform__start_league_quantile_uses_existing_player_ratings(base_cn):
+    """
+    Bug reproduction: start_league_quantile should use percentile of existing player
+    ratings for new players, but update_players_to_leagues is never called so
+    _league_player_ratings stays empty and all new players get default rating.
+
+    Expected: New player P_NEW should start at 5th percentile of existing ratings (~920)
+    Actual: New player starts at default 1000 because _league_player_ratings is empty
+    """
+    import numpy as np
+
+    num_existing_players = 60
+    player_ids = [f"P{i}" for i in range(num_existing_players)]
+    team_ids = [f"T{i % 2 + 1}" for i in range(num_existing_players)]
+
+    df1 = pl.DataFrame(
+        {
+            "pid": player_ids,
+            "tid": team_ids,
+            "mid": ["M1"] * num_existing_players,
+            "dt": ["2024-01-01"] * num_existing_players,
+            "perf": [0.3 + (i % 10) * 0.07 for i in range(num_existing_players)],
+            "pw": [1.0] * num_existing_players,
+        }
+    )
+
+    gen = PlayerRatingGenerator(
+        performance_column="perf",
+        column_names=base_cn,
+        auto_scale_performance=True,
+        start_league_quantile=0.05,
+        start_min_count_for_percentiles=50,
+        features_out=[RatingKnownFeatures.PLAYER_OFF_RATING],
+    )
+    gen.fit_transform(df1)
+
+    existing_ratings = [
+        gen._player_off_ratings[pid].rating_value for pid in player_ids
+    ]
+    expected_quantile_rating = np.percentile(existing_ratings, 5)
+
+    srg = gen.start_rating_generator
+    assert len(srg._league_player_ratings.get(None, [])) >= 50, (
+        f"Expected _league_player_ratings to have >=50 entries but got "
+        f"{len(srg._league_player_ratings.get(None, []))}. "
+        "update_players_to_leagues is never called."
+    )
+
+    df2 = pl.DataFrame(
+        {
+            "pid": ["P_NEW", "P0"],
+            "tid": ["T1", "T2"],
+            "mid": ["M2", "M2"],
+            "dt": ["2024-01-02", "2024-01-02"],
+            "pw": [1.0, 1.0],
+        }
+    )
+    result = gen.future_transform(df2)
+
+    new_player_start_rating = result.filter(pl.col("pid") == "P_NEW")[
+        "player_off_rating_perf"
+    ][0]
+
+    assert new_player_start_rating == pytest.approx(expected_quantile_rating, rel=0.1), (
+        f"New player should start at 5th percentile ({expected_quantile_rating:.1f}) "
+        f"but got {new_player_start_rating:.1f}. "
+        "start_league_quantile has no effect because update_players_to_leagues is never called."
+    )