Fix data pipelines once and for all

pipeline/transform/price_estimation/utils.py

@@ -36,6 +36,20 @@ SHRINKAGE_K = 50
 # noisy year) without flattening genuine multi-year trends.
 TEMPORAL_SMOOTHNESS_LAMBDA = 0.05
 
+# Per-year support scaling for the temporal smoothness penalty. A flat lambda
+# is too weak for years with very few repeat-sale pairs: a sector can have
+# hundreds of pairs overall (so cell-level n/(n+k) shrinkage barely moves it)
+# yet have individual years estimated from 1-2 pairs, producing 2-7x
+# single-year index spikes. Each curvature row is therefore scaled by the
+# local pair support of its year triple:
+#   lambda_eff = lambda0 * (1 + SMOOTHNESS_SUPPORT_PAIRS / s)
+# where s is the minimum cross-year pair count among the triple's years.
+# Well-supported years (s >> SMOOTHNESS_SUPPORT_PAIRS) keep lambda_eff ~
+# lambda0 (current behaviour); a year identified by a single pair gets
+# ~41x lambda0, pulling its beta strongly toward the local trend through its
+# neighbours. Same-year pairs cancel in the design and are not counted.
+SMOOTHNESS_SUPPORT_PAIRS = 40
+
 
 def type_group_expr():
     """Polars expression: Property type -> type_group."""