spb/wp14_uqo

1% ============================================================================2\section{Introduction}3\label{sec:introduction}4% ============================================================================56The hedonic price framework decomposes the market value of a heterogeneous good7into the implicit prices of its characteristics \citep{lancaster1966new,8rosen1974hedonic}. In housing markets it remains the workhorse tool for appraisal,9index construction, property-tax assessment, and policy analysis. Yet the empirical10performance of hedonic models is bounded by the richness of the characteristics an11analyst can measure. A large share of what determines a home's value---its state of12repair, the quality of its finishes, its curb appeal, the character of its13street---is recorded not in structured fields but in the free-text description and14the photographs that accompany each listing. These modalities have long been15discarded because they were costly to encode at scale, leaving hedonic estimates16exposed to omitted-quality bias and hedonic predictions short of their attainable17accuracy.1819This paper treats general-purpose artificial-intelligence models as a20\emph{feature-engineering technology} for hedonic analysis and quantifies the21explanatory power that technology adds. We assemble a cross-section of 19{,}12322single-family houses and townhouses listed for sale in Quebec on 3~July~2026,23pooling the brokerage-listed MLS inventory from Realtor.ca with the24for-sale-by-owner inventory from DuProprio. On top of standard structural,25locational, and amenity attributes, we build two machine-generated feature layers.26First, a large language model (LLM) reads each free-text description and returns a27fixed set of property specifications---state of repair, renovation of the kitchen28and bathroom, flooring and countertop materials, heating system, view, and income29potential---under a strict JSON schema with closed enumerations. Second, a30vision--language model inspects the façade photograph of each property and returns31visual attributes: overall condition, curb appeal, renovation level, exterior32materials, landscaping, setting, and a luxury indicator. Because both extractions33are schema-constrained, their outputs enter the hedonic regression as ordinary34tabulated regressors with interpretable coefficients rather than as latent scores.3536Our contribution is threefold. First, we show that AI-based feature engineering37delivers gains in explanatory power over an already rich hedonic specification. In38our preferred specification---a log-linear model with 195 municipality fixed39effects, property-type fixed effects, and fifteen keyword amenities, room-level and40listing-quality controls---the two machine-generated layers raise in-sample $R^2$41from 0.717 to 0.755, and they raise the out-of-sample $R^2$ of every estimator we42consider, from ordinary least squares to gradient boosting and an embedding-based43neural network. Of the two layers, the façade-vision layer is by far the more44informative: it adds 3.0 points of in-sample $R^2$ against 0.8 for the text layer,45and 5.8 (2.7) points of out-of-sample $R^2$ for OLS (gradient boosting), more than46the gain from replacing OLS by a random forest. Second, we show that the extracted47attributes carry interpretable implicit prices---a one-point increase in visual curb48appeal is associated with a 2.3\% higher asking price, a visible luxury cue with495.2\%, brick or stone façades with premia of 3--4\% relative to the omitted50materials, and a renovated kitchen (read from the text) with 2.6\%---and that these51associations survive municipality-clustered inference, a cluster bootstrap, finer52postal-code (RTA) fixed effects, and a control for the language of the description.53Third, we document a complete and reproducible pipeline from raw scraped listings to54an analysis-ready dataset, including rules-based outlier cleaning, unit harmonization55that reconstructs living area from room-level dimensions, and the56schema-constrained extraction protocol, with a total AI cost of about \$60 for the57full sample---roughly a third of a cent per listing.5859The revision reported here (version 2.1) also confronts three threats that a60skeptical reader should raise against any ``AI features'' exercise, and we61believe the paper is more useful for addressing them explicitly. The first is62\emph{information leakage}. Every DuProprio description in our data embeds a63structured block that includes the asking price itself (``Asking Price \$599,000''),64so the language model read the dependent variable for 19.7\% of the sample. The65schema contains no price field, but the text-layer gain is five times larger on66DuProprio than on Realtor.ca, whose descriptions never contain the asking price,67and we therefore report all gains separately by platform and treat the Realtor.ca68estimates as the clean benchmark. The vision model never sees the price. The second69is \emph{language}. Forty-three percent of the descriptions are in English, and70English-language listings are priced about 7\% higher than French ones conditional71on all other attributes and on location; the AI implicit prices are, however,72essentially unchanged when language is controlled for, and the gains from both73layers are present within the French and within the English subsamples. The third74is \emph{spatial dependence in validation}. Random cross-validation folds let a75model learn from a neighbour of each held-out house; when folds are formed by76forward sortation area instead, all estimators lose three to five points of77out-of-sample $R^2$, but the increments from the text and vision layers are78preserved.7980Our main findings can be summarized as follows. The preferred specification explains8175.5\% of the variation in log asking prices when living area is observed and 72.5\%82on the full sample when it is not. The living-area elasticity is 0.36, each83additional bathroom is associated with a 7.6\% premium, and the age profile traces84the familiar depreciation-then-heritage pattern. The municipality effects range from85about $+231$\% in Westmount to about $-30$\% in remote municipalities relative to the86pooled group of small municipalities, and order the province's markets as expected.87On the identical feature set, an equally weighted blend of a robust embedding-based88neural network and gradient boosting attains an out-of-sample $R^2$ of 0.779 with a89median absolute percentage error (MdAPE) of 14.0\% under random folds. We interpret90all coefficients as \emph{predictive associations} rather than causal effects: listing91content is chosen by sellers and their agents and correlates with unobserved92determinants of value, and the dependent variable is an asking price, two issues we93discuss at length in Section~\ref{sec:robustness}.9495Beyond the Quebec application, the exercise speaks to a measurement problem in96housing economics. Text- and image-derived quality measures have until now required97bespoke supervised models trained on labelled data98\citep{nowak2017textual,shen2021information,johnson2020curb,lindenthal2021architecture}.99General-purpose foundation models with schema-constrained decoding lower the fixed100cost of that measurement step to nearly zero, produce attributes that are101human-readable rather than latent, and can be audited field by field. Our results102suggest that this technology shifts the effective characteristics frontier of hedonic103analysis by enough to matter for appraisal, automated valuation, and assessment104applications---provided that the text fed to the model is screened for the outcome105it is meant to predict.106107The remainder of the paper is organized as follows. Section~\ref{sec:literature}108reviews the related literature. Section~\ref{sec:data} describes the data, the109cleaning rules, and the content of the descriptions. Section~\ref{sec:methodology}110presents the feature-engineering pipeline and the econometric specification.111Section~\ref{sec:results} reports the main results. Section~\ref{sec:benchmark}112maps the predictive frontier across estimators. Section~\ref{sec:robustness}113reports the robustness analyses and discusses limitations.114Section~\ref{sec:conclusion} concludes.115