spb/wp11_uqo Public

1% Author: Simon-Pierre Boucher — contact@spboucher.ai2% ============================================================================3\section{Related literature}4\label{sec:lit}56\subsection{The functional-form debate}78Hedonic theory is silent about functional form. \citet{rosen1974hedonic}9showed that the price surface is an equilibrium envelope of bids and10offers, so nothing structural pins down its curvature ---11form is an empirical choice. The classical exchange settled into a12three-cornered debate: \citet{halvorsen1981choice} advocated data-driven13Box--Cox transformations \citep{box1964analysis};14\citet{cassel1985cost} cautioned that flexible transformations optimize15in-sample fit while degrading the implicit prices researchers actually16use; and \citet{cropper1988choice}, in an influential simulation study,17found that simple forms --- linear and semi-log --- are the most robust to18the misspecification and omitted variables that characterize real data.19Fifty years on, the semi-log remains the profession's default20\citep{malpezzi2003hedonic, sirmans2005composition}, but systematic21head-to-head evidence on modern samples remains scarce, and the classic22studies predate the out-of-sample validation culture. Two further23technical strands feed our design: retransformation bias in log models24\citep{duan1983smearing}, which we correct with the smearing factor so25that every form competes in levels; and forecast-evaluation methodology26for house prices \citep{clapp2002predicting, steurer2021metrics}.2728\subsection{Space and time in hedonic models}2930A second literature concerns what the covariates cannot capture. House31prices are spatially autocorrelated at fine scale32\citep{dubin1998spatial, anselin1988spatial}, and the practical remedies33range from submarket segmentation \citep{goodman1998housing} to spatial34econometrics and simple location fixed effects; comparative studies35generally find that flexibly absorbing location buys more predictive36accuracy than any parametric spatial process \citep{case2004modeling,37bourassa2010predicting, pace1997spatial}. In the Quebec context,38\citet{desrosiers2000hedonic} document the importance of access and39neighbourhood attributes in Quebec City. On the time dimension, hedonic40price-index theory --- the time-dummy versus imputation debate ---41is surveyed by \citet{hill2013hedonic} and codified in the RPPI handbook42\citep{oecd2013handbook, silver2018house}; a byproduct of our month-FE43models is a direct test of how sensitive a constant-quality index is to44the underlying form.4546\subsection{Machine learning and automated valuation}4748The third strand is the rapid colonization of mass valuation by machine49learning. \citet{mullainathan2017machine} frame prediction as ML's50comparative advantage, with hedonic pricing as their running example;51\citet{athey2019machine} survey the toolkit. Applied comparisons ---52random forests \citep{breiman2001random}, gradient boosting53\citep{friedman2001greedy, ke2017lightgbm} --- consistently report 20--40\%54error reductions over linear hedonics \citep{mayer2019estimation,55hong2020machine, kok2017big}, with accuracy degrading in thin rural56markets \citep{bogin2020house}; \citet{gencay1996nonlinear} anticipated57the pattern with semiparametric estimators. Three gaps motivate our58design. First, most comparisons change several ingredients at once59(features, sample, validation scheme), so the \emph{sources} of the ML60advantage --- curvature? interactions? spatial flexibility? --- are rarely61isolated; our axis design holds features constant and varies one choice62at a time. Second, almost all published comparisons use random63cross-validation, which leaks future price levels into training and64flatters any flexible model; we score every model under a65forward-in-time split as well. Third, comparisons seldom examine what the66flexible models \emph{imply} --- indices, implicit prices --- which is67what economists need; our extensions section does exactly that.68