idk

pipeline/transform/price_estimation/backtest.py

@@ -11,9 +11,9 @@ from pathlib import Path
 import numpy as np
 import polars as pl
 
+from pipeline.transform.price_estimation.estimate import guarded_blend_estimates
 from pipeline.transform.price_estimation.index import build_index
 from pipeline.transform.price_estimation.knn import (
-    KNN_BLEND_WEIGHT,
     build_knn_pool,
     knn_median_psm,
 )
@@ -115,7 +115,10 @@ def predict(test: pl.DataFrame, index: pl.DataFrame) -> pl.DataFrame:
             .clip(-MAX_LOG_ADJUSTMENT, MAX_LOG_ADJUSTMENT)
             .exp()
         )
-        .fill_null(pl.col("input_price").cast(pl.Float64))
+        # Keep null when the index can't be interpolated, matching production
+        # (estimate.py ships null there). compute_metrics filters to finite
+        # positive predictions, so these rows correctly drop from the Index n
+        # rather than silently degrading to the Naive prediction.
         .alias("predicted"),
     )
     return test
@@ -265,13 +268,12 @@ def main():
         f"  kNN estimates: {n_knn:,} of {len(test):,} ({n_knn / len(test) * 100:.1f}%)"
     )
 
-    # Blend: (1-w)*index + w*kNN where both available
+    # Blend with the exact shipped estimator (stability gate + last-price cap +
+    # null-when-no-index) so the "Blended" stage reflects production accuracy.
+    # input_price is the backtest equivalent of production's "Last known price".
     index_est = test["predicted"].to_numpy().astype(np.float64)
-    knn_valid = np.isfinite(knn_est) & (knn_est > 0)
-    blended = np.where(
-        knn_valid & np.isfinite(index_est),
-        (1 - KNN_BLEND_WEIGHT) * index_est + KNN_BLEND_WEIGHT * knn_est,
-        np.where(np.isfinite(index_est), index_est, knn_est),
+    blended = guarded_blend_estimates(
+        index_est, knn_est, test["input_price"].cast(pl.Float64).to_numpy()
     )
 
     actual = test["actual_price"].to_numpy().astype(np.float64)