spb/wp11_uqo Public

1% Author: Simon-Pierre Boucher — contact@spboucher.ai2% ============================================================================3\section{Conclusion}4\label{sec:concl}56We ran the hedonic specification debate as a controlled experiment:7twenty models, four design axes, one attribute set, one scoreboard, two8holdouts, half a million sales. The verdict inverts the profession's9implicit priorities. The choices that consume the most methodological10attention --- functional form, transformation parameters --- move11out-of-sample accuracy by at most two points of median error, while the12choices often made by default --- whether and how finely to absorb space13and time --- move it by three to ten. Machine-learning estimators14dominate the standard scoreboard, but the standard scoreboard asks the15wrong question: when the test set lies in the future, as it always does16in deployment, the boosted ensemble's advantage reverses against a17spline hedonic with good controls, because most of what it learned was a18spatially fine price configuration that does not transfer across the19market cycle.2021Meanwhile the objects economists actually harvest from hedonic models22prove reassuringly stable: constant-quality indices agree across forms23to within a few points over a full boom--bust--recovery cycle, and the24machine's implicit age and area gradients replicate the quadratic OLS.25The practical synthesis is almost embarrassingly simple: a semi-log with26splines, quarter effects, and spatial fixed effects at a granularity27matched to market thickness is within two points of the frontier in28deployment conditions --- transparent, auditable, and fast. Where the29extra accuracy of learning methods is worth its retraining cadence ---30high-frequency AVMs, collateral monitoring --- our decomposition says to31spend the flexibility budget on the spatial surface, and to validate,32always, against the future rather than against a shuffle of the past.33