High-value stock and the pull to the mean.

High-value stock deserves its own validation cohort. Measure the error on your completed cases before an estimate changes a lending route.

A valuation model can make its costliest errors in pounds on the most expensive homes. Sparse comparables and unusual features make a single market-wide score a poor substitute for testing the segment itself.

Why the top of the market breaks models

High-value stock is, almost by definition, unusual. Fewer comparables, more idiosyncratic features, wider genuine dispersion in what similar-looking homes achieve. A model trained to minimise error across the whole market learns that the safest guess for an outlier is to pull it toward the mean.

That pull can disappear inside a national score. The properties where it matters carry the largest exposures, so even a modest percentage error can represent a large number of pounds.

Reverting to the area average is where a lazy model costs the most, quietly, on the loans that are largest.

Segment the test before trusting the score

Pretira treats premium stock as a separate evaluation segment because a national aggregate can hide the error distribution that matters to a high-value book. We do not ask a lender to accept a lift claim on trust. Use a bounded evaluation to measure the estimate, range and routing signal against your own completed cases.

Route it like the risk it is

IF property_value >= 1000000 AND property_value_confidence < HIGH -> desktop or physical val

This is an illustrative routing pattern, not a lending policy. The point is to test whether the estimate, range and confidence band separate routine cases from exceptions on your own high-value cohort. Only then should those signals influence which cases stay on data and which go to a desktop or physical valuation.