Most mass-appraisal models are applied in a single step: the calibrated model writes the Just Values directly to the roll. In that environment, holdout testing isn't optional—it is the primary evidence that the model generalizes and that residual independence holds.
Traditional
theoretical tests such as Durbin-Watson are rarely informative with large
cross-sectional sales files. A properly constructed forward (out-of-time)
holdout sample is far more useful: it shows whether the model’s accuracy, ratio
performance, and residual behavior remain stable on sales the model has never
seen.
Instead of relying on random in-sample 80/20 splits that share the
same historical noise, we break down how to construct a true out-of-time
forward sales holdout sample (July–Oct 2026), handle missing cadastral data
through strict field-audit protocols rather than flawed medians, and execute
bidirectional time adjustments to maintain absolute valuation parity.
Session
8 of the Mass Appraisal Modeling series walks through the full holdout protocol—sample construction, representativeness checks, pre- and post-outlier ratio
comparison, and Just Value ratio analysis—and shows why this practical
test is the right way to evaluate residual independence in modern CAMA work.
Live
this Wednesday.
Executive Summary is free for everyone.
No comments:
Post a Comment