spb/wp11_uqo Public

1% Author: Simon-Pierre Boucher — contact@spboucher.ai2% ============================================================================3\section{Discussion}4\label{sec:disc}56\subsection{A hierarchy of specification choices}78Read jointly, the four axes imply a clear hierarchy of returns to9specification effort, useful to anyone building a hedonic model:1011\begin{enumerate}12\item \textbf{Get location right} ($\sim$10 points of median error):13any spatial absorption beats none by a margin that dwarfs every other14choice; the optimal granularity is interior --- around 5~km here ---15because finer cells trade bias for variance.16\item \textbf{Include some time control} ($\sim$3 points): in a moving17market, a model without time effects mistakes appreciation for18attributes; granularity beyond quarters is cosmetic.19\item \textbf{Choose the estimator for the job} ($\sim$2 points20deployed, $\sim$4 under random validation): boosting buys real accuracy21in interpolation-heavy uses (filling gaps within a period, mass22appraisal with contemporaneous comparables) and much less when the23target is the future.24\item \textbf{Stop worrying about functional form} ($\lesssim$225points): splines are a cheap upgrade; the Box--Cox machinery is not26worth its complexity, exactly as \citet{cropper1988choice} concluded27from simulations four decades ago.28\end{enumerate}2930\subsection{Why the machines lose their edge out of time}3132The decomposition rows explain the generalization gap. The boosting33model's random-split advantage comes almost entirely from a flexible34spatial surface (removing coordinates costs it 8.8 points). That surface35is estimated \emph{jointly} with the time path, and trees clamp: beyond36the last training month the model prices 2025--26 sales at a frozen37level, exactly like the carry-forward linear models, while its38fine-grained spatial fit --- calibrated on the 2021--24 price39configuration --- partially decays as relative prices shift. The linear40models, coarser but more rigid, carry a structure that transfers better.41This is not an indictment of machine learning --- retrained monthly, the42boosting model would presumably keep its interpolation advantage --- but43it prices the \emph{retraining requirement} that random cross-validation44hides, and it matches the practitioner evidence that AVM accuracy decays45quickly out of sample period \citep{bogin2020house, kok2017big}.4647\subsection{Implications}4849\paragraph{For research.} Hedonic coefficients and indices are robust50objects: form and even estimator perturb them little51(Section~\ref{sec:ext}). Researchers using hedonics to \emph{measure}52--- indices, implicit prices, capitalization effects --- can keep simple53forms with good controls and report validation under a temporal split54when prediction claims are made.5556\paragraph{For mass appraisal.} Quebec's assessors face exactly this57design problem every three years. Our results suggest the largest58accuracy gains lie not in exotic estimators but in spatial resolution59chosen by market thickness --- and that any model, linear or boosted,60frozen at a reference date degrades by 3--7 points of median error61within two years. This quantifies the staleness mechanism behind the62assessment inequities documented in our companion paper (UQO WP10): the63valuation technology's out-of-time decay is of the same magnitude as the64inequities measured there.6566\paragraph{For the AVM literature.} Every comparison should report a67forward-in-time split alongside random cross-validation68\citep{steurer2021metrics}. On our data, the choice of split changes not69just magnitudes but the \emph{winner}.7071\subsection{Limitations}7273Our attribute set, though assessor-grade, omits listing-level quality74(renovations, finishes, views); richer features would likely raise the75machines' interpolation advantage. Hyper-parameters were deliberately76conventional --- a tuned boosted model would gain somewhat, as would a77spatially smarter linear model (e.g.\ kriging residuals,78\citealp{case2004modeling}). The forward split covers one specific79regime change (the 2025--26 recovery); other windows would shift80magnitudes. And Quebec's institutional homogeneity aids transferability81of results across municipalities but may limit it across countries.82