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

\subsection{The functional-form debate}

Hedonic theory is silent about functional form. \citet{rosen1974hedonic}
showed that the price surface is an equilibrium envelope of bids and
offers, so nothing structural pins down its curvature ---
form is an empirical choice. The classical exchange settled into a
three-cornered debate: \citet{halvorsen1981choice} advocated data-driven
Box--Cox transformations \citep{box1964analysis};
\citet{cassel1985cost} cautioned that flexible transformations optimize
in-sample fit while degrading the implicit prices researchers actually
use; and \citet{cropper1988choice}, in an influential simulation study,
found that simple forms --- linear and semi-log --- are the most robust to
the misspecification and omitted variables that characterize real data.
Fifty years on, the semi-log remains the profession's default
\citep{malpezzi2003hedonic, sirmans2005composition}, but systematic
head-to-head evidence on modern samples remains scarce, and the classic
studies predate the out-of-sample validation culture. Two further
technical strands feed our design: retransformation bias in log models
\citep{duan1983smearing}, which we correct with the smearing factor so
that every form competes in levels; and forecast-evaluation methodology
for house prices \citep{clapp2002predicting, steurer2021metrics}.

\subsection{Space and time in hedonic models}

A second literature concerns what the covariates cannot capture. House
prices are spatially autocorrelated at fine scale
\citep{dubin1998spatial, anselin1988spatial}, and the practical remedies
range from submarket segmentation \citep{goodman1998housing} to spatial
econometrics and simple location fixed effects; comparative studies
generally find that flexibly absorbing location buys more predictive
accuracy than any parametric spatial process \citep{case2004modeling,
bourassa2010predicting, pace1997spatial}. In the Quebec context,
\citet{desrosiers2000hedonic} document the importance of access and
neighbourhood attributes in Quebec City. On the time dimension, hedonic
price-index theory --- the time-dummy versus imputation debate ---
is surveyed by \citet{hill2013hedonic} and codified in the RPPI handbook
\citep{oecd2013handbook, silver2018house}; a byproduct of our month-FE
models is a direct test of how sensitive a constant-quality index is to
the underlying form.

\subsection{Machine learning and automated valuation}

The third strand is the rapid colonization of mass valuation by machine
learning. \citet{mullainathan2017machine} frame prediction as ML's
comparative advantage, with hedonic pricing as their running example;
\citet{athey2019machine} survey the toolkit. Applied comparisons ---
random forests \citep{breiman2001random}, gradient boosting
\citep{friedman2001greedy, ke2017lightgbm} --- consistently report 20--40\%
error reductions over linear hedonics \citep{mayer2019estimation,
hong2020machine, kok2017big}, with accuracy degrading in thin rural
markets \citep{bogin2020house}; \citet{gencay1996nonlinear} anticipated
the pattern with semiparametric estimators. Three gaps motivate our
design. First, most comparisons change several ingredients at once
(features, sample, validation scheme), so the \emph{sources} of the ML
advantage --- curvature? interactions? spatial flexibility? --- are rarely
isolated; our axis design holds features constant and varies one choice
at a time. Second, almost all published comparisons use random
cross-validation, which leaks future price levels into training and
flatters any flexible model; we score every model under a
forward-in-time split as well. Third, comparisons seldom examine what the
flexible models \emph{imply} --- indices, implicit prices --- which is
what economists need; our extensions section does exactly that.