% Author: Simon-Pierre Boucher — contact@spboucher.ai
% ============================================================================
\section{Conclusion}
\label{sec:concl}

We ran the hedonic specification debate as a controlled experiment:
twenty models, four design axes, one attribute set, one scoreboard, two
holdouts, half a million sales. The verdict inverts the profession's
implicit priorities. The choices that consume the most methodological
attention --- functional form, transformation parameters --- move
out-of-sample accuracy by at most two points of median error, while the
choices often made by default --- whether and how finely to absorb space
and time --- move it by three to ten. Machine-learning estimators
dominate the standard scoreboard, but the standard scoreboard asks the
wrong question: when the test set lies in the future, as it always does
in deployment, the boosted ensemble's advantage reverses against a
spline hedonic with good controls, because most of what it learned was a
spatially fine price configuration that does not transfer across the
market cycle.

Meanwhile the objects economists actually harvest from hedonic models
prove reassuringly stable: constant-quality indices agree across forms
to within a few points over a full boom--bust--recovery cycle, and the
machine's implicit age and area gradients replicate the quadratic OLS.
The practical synthesis is almost embarrassingly simple: a semi-log with
splines, quarter effects, and spatial fixed effects at a granularity
matched to market thickness is within two points of the frontier in
deployment conditions --- transparent, auditable, and fast. Where the
extra accuracy of learning methods is worth its retraining cadence ---
high-frequency AVMs, collateral monitoring --- our decomposition says to
spend the flexibility budget on the spatial surface, and to validate,
always, against the future rather than against a shuffle of the past.