From: Jérôme Benoit <jerome.benoit@piment-noir.org>
Date: Sun, 24 May 2026 00:06:18 +0000 (+0200)
Subject: feat(weights): compose per-label weights with temporal decay before model.fit
X-Git-Url: https://git.piment-noir.org/?a=commitdiff_plain;h=635a3b9e06d1e69ceea80cd4199b6b6453c0ffa1;p=freqai-strategies.git

feat(weights): compose per-label weights with temporal decay before model.fit
---

diff --git a/quickadapter/user_data/freqaimodels/QuickAdapterRegressorV3.py b/quickadapter/user_data/freqaimodels/QuickAdapterRegressorV3.py
index 661bce1..99c3b06 100644
--- a/quickadapter/user_data/freqaimodels/QuickAdapterRegressorV3.py
+++ b/quickadapter/user_data/freqaimodels/QuickAdapterRegressorV3.py
@@ -50,6 +50,7 @@ from Utils import (
     LABEL_COLUMNS,
     REGRESSORS,
     Regressor,
+    compose_sample_weights,
     ensure_datetime_series,
     eval_set_and_weights,
     fit_regressor,
@@ -114,6 +115,40 @@ class QuickAdapterRegressorV3(BaseRegressionModel):
             if not c.endswith(self._LABEL_WEIGHT_SUFFIX)
         ]
 
+    def _compose_train_weights(
+        self,
+        dd: dict[str, Any],
+        unfiltered_df: pd.DataFrame,
+        dk: FreqaiDataKitchen,
+    ) -> dict[str, Any]:
+        weight_cols = {
+            label: f"{label}{self._LABEL_WEIGHT_SUFFIX}"
+            for label in dk.label_list
+            if f"{label}{self._LABEL_WEIGHT_SUFFIX}" in unfiltered_df.columns
+        }
+        if not weight_cols:
+            return dd
+        label_weights_map: dict[str, NDArray] = {}
+        for label, weight_col in weight_cols.items():
+            series = unfiltered_df[weight_col]
+            train_w = series.reindex(dd["train_features"].index, fill_value=1.0).to_numpy(dtype=float)
+            label_weights_map[label] = train_w
+        dd["train_weights"] = compose_sample_weights(
+            np.asarray(dd["train_weights"], dtype=float),
+            label_weights_map,
+        )
+        if dd.get("test_features") is not None and len(dd["test_features"]) > 0:
+            test_map: dict[str, NDArray] = {}
+            for label, weight_col in weight_cols.items():
+                series = unfiltered_df[weight_col]
+                test_w = series.reindex(dd["test_features"].index, fill_value=1.0).to_numpy(dtype=float)
+                test_map[label] = test_w
+            dd["test_weights"] = compose_sample_weights(
+                np.asarray(dd["test_weights"], dtype=float),
+                test_map,
+            )
+        return dd
+
     _SQRT_2: Final[float] = np.sqrt(2.0)
 
     _OPTUNA_LABEL_N_OBJECTIVES: Final[int] = 7