% ============================================================================
\section{Introduction}
\label{sec:introduction}
% ============================================================================

The hedonic price framework decomposes the market value of a heterogeneous good
into the implicit prices of its characteristics \citep{lancaster1966new,
rosen1974hedonic}. In housing markets it remains the workhorse tool for appraisal,
index construction, property-tax assessment, and policy analysis. Yet the empirical
performance of hedonic models is bounded by the richness of the characteristics an
analyst can measure. A large share of what determines a home's value---its state of
repair, the quality of its finishes, its curb appeal, the character of its
street---is recorded not in structured fields but in the free-text description and
the photographs that accompany each listing. These modalities have long been
discarded because they were costly to encode at scale, leaving hedonic estimates
exposed to omitted-quality bias and hedonic predictions short of their attainable
accuracy.

This paper treats general-purpose artificial-intelligence models as a
\emph{feature-engineering technology} for hedonic analysis and quantifies the
explanatory power that technology adds. We assemble a cross-section of 19{,}123
single-family houses and townhouses listed for sale in Quebec on 3~July~2026,
pooling the brokerage-listed MLS inventory from Realtor.ca with the
for-sale-by-owner inventory from DuProprio. On top of standard structural,
locational, and amenity attributes, we build two machine-generated feature layers.
First, a large language model (LLM) reads each free-text description and returns a
fixed set of property specifications---state of repair, renovation of the kitchen
and bathroom, flooring and countertop materials, heating system, view, and income
potential---under a strict JSON schema with closed enumerations. Second, a
vision--language model inspects the façade photograph of each property and returns
visual attributes: overall condition, curb appeal, renovation level, exterior
materials, landscaping, setting, and a luxury indicator. Because both extractions
are schema-constrained, their outputs enter the hedonic regression as ordinary
tabulated regressors with interpretable coefficients rather than as latent scores.

Our contribution is threefold. First, we show that AI-based feature engineering
delivers gains in explanatory power over an already rich hedonic specification. In
our preferred specification---a log-linear model with 195 municipality fixed
effects, property-type fixed effects, and fifteen keyword amenities, room-level and
listing-quality controls---the two machine-generated layers raise in-sample $R^2$
from 0.717 to 0.755, and they raise the out-of-sample $R^2$ of every estimator we
consider, from ordinary least squares to gradient boosting and an embedding-based
neural network. Of the two layers, the façade-vision layer is by far the more
informative: it adds 3.0 points of in-sample $R^2$ against 0.8 for the text layer,
and 5.8 (2.7) points of out-of-sample $R^2$ for OLS (gradient boosting), more than
the gain from replacing OLS by a random forest. Second, we show that the extracted
attributes carry interpretable implicit prices---a one-point increase in visual curb
appeal is associated with a 2.3\% higher asking price, a visible luxury cue with
5.2\%, brick or stone façades with premia of 3--4\% relative to the omitted
materials, and a renovated kitchen (read from the text) with 2.6\%---and that these
associations survive municipality-clustered inference, a cluster bootstrap, finer
postal-code (RTA) fixed effects, and a control for the language of the description.
Third, we document a complete and reproducible pipeline from raw scraped listings to
an analysis-ready dataset, including rules-based outlier cleaning, unit harmonization
that reconstructs living area from room-level dimensions, and the
schema-constrained extraction protocol, with a total AI cost of about \$60 for the
full sample---roughly a third of a cent per listing.

The revision reported here (version 2.1) also confronts three threats that a
skeptical reader should raise against any ``AI features'' exercise, and we
believe the paper is more useful for addressing them explicitly. The first is
\emph{information leakage}. Every DuProprio description in our data embeds a
structured block that includes the asking price itself (``Asking Price \$599,000''),
so the language model read the dependent variable for 19.7\% of the sample. The
schema contains no price field, but the text-layer gain is five times larger on
DuProprio than on Realtor.ca, whose descriptions never contain the asking price,
and we therefore report all gains separately by platform and treat the Realtor.ca
estimates as the clean benchmark. The vision model never sees the price. The second
is \emph{language}. Forty-three percent of the descriptions are in English, and
English-language listings are priced about 7\% higher than French ones conditional
on all other attributes and on location; the AI implicit prices are, however,
essentially unchanged when language is controlled for, and the gains from both
layers are present within the French and within the English subsamples. The third
is \emph{spatial dependence in validation}. Random cross-validation folds let a
model learn from a neighbour of each held-out house; when folds are formed by
forward sortation area instead, all estimators lose three to five points of
out-of-sample $R^2$, but the increments from the text and vision layers are
preserved.

Our main findings can be summarized as follows. The preferred specification explains
75.5\% of the variation in log asking prices when living area is observed and 72.5\%
on the full sample when it is not. The living-area elasticity is 0.36, each
additional bathroom is associated with a 7.6\% premium, and the age profile traces
the familiar depreciation-then-heritage pattern. The municipality effects range from
about $+231$\% in Westmount to about $-30$\% in remote municipalities relative to the
pooled group of small municipalities, and order the province's markets as expected.
On the identical feature set, an equally weighted blend of a robust embedding-based
neural network and gradient boosting attains an out-of-sample $R^2$ of 0.779 with a
median absolute percentage error (MdAPE) of 14.0\% under random folds. We interpret
all coefficients as \emph{predictive associations} rather than causal effects: listing
content is chosen by sellers and their agents and correlates with unobserved
determinants of value, and the dependent variable is an asking price, two issues we
discuss at length in Section~\ref{sec:robustness}.

Beyond the Quebec application, the exercise speaks to a measurement problem in
housing economics. Text- and image-derived quality measures have until now required
bespoke supervised models trained on labelled data
\citep{nowak2017textual,shen2021information,johnson2020curb,lindenthal2021architecture}.
General-purpose foundation models with schema-constrained decoding lower the fixed
cost of that measurement step to nearly zero, produce attributes that are
human-readable rather than latent, and can be audited field by field. Our results
suggest that this technology shifts the effective characteristics frontier of hedonic
analysis by enough to matter for appraisal, automated valuation, and assessment
applications---provided that the text fed to the model is screened for the outcome
it is meant to predict.

The remainder of the paper is organized as follows. Section~\ref{sec:literature}
reviews the related literature. Section~\ref{sec:data} describes the data, the
cleaning rules, and the content of the descriptions. Section~\ref{sec:methodology}
presents the feature-engineering pipeline and the econometric specification.
Section~\ref{sec:results} reports the main results. Section~\ref{sec:benchmark}
maps the predictive frontier across estimators. Section~\ref{sec:robustness}
reports the robustness analyses and discusses limitations.
Section~\ref{sec:conclusion} concludes.