Python 52.4%
TeX 47.6%
1% ============================================================================2\section{Related Literature}3\label{sec:literature}4% ============================================================================56This paper connects four strands of research.78The first is the classical \textbf{hedonic pricing} literature. Building on9\citet{lancaster1966new}, \citet{rosen1974hedonic} formalized the market for10differentiated goods as a market for their underlying characteristics, providing the11theoretical foundation for interpreting regression coefficients as implicit prices.12The empirical housing literature that followed is vast; \citet{sirmans2005composition}13survey the characteristics most commonly included in hedonic house-price models and14document substantial heterogeneity in specification, \citet{malpezzi2003hedonic}15provides a critical review of functional-form and identification choices, and16\citet{hill2013hedonic} surveys the use of hedonic methods for house-price17measurement. A recurring theme is omitted-variable bias from unmeasured quality:18the state of repair, the standard of finishes, and the visual quality of a dwelling19are priced by the market but rarely recorded in the structured fields available to20the econometrician. \citet{nowak2020quality} show that this omission is not21innocuous even for aggregate statistics, since quality-adjusted repeat-sales indexes22built with information from listing remarks differ materially from conventional23ones. Our contribution speaks directly to this measurement problem. Our log-linear24specification, with logged continuous attributes and a dependent variable in logs,25follows the mainstream of this literature, and we read indicator coefficients with26the \citet{halvorsen1980interpretation} transformation; the small-sample refinement27of \citet{kennedy1981estimation} is immaterial at our precision.2829The second strand concerns \textbf{spatial structure, inference, and prediction}.30\citet{can1992spatial} and \citet{anselin1988spatial} emphasize that unobserved31locational amenities induce spatial dependence that, if ignored, biases coefficient32estimates and understates uncertainty; we absorb the first-order component of this33problem through municipality fixed effects, re-estimate with finer forward34sortation area (RTA) effects, and cluster inference at the level of the fixed35effects \citep{cameron2015cluster,abadie2023when}, complementing the analytical36standard errors with a cluster bootstrap because our clusters are of very unequal37size \citep{mackinnon2017wild}. The same dependence bears on validation.38\citet{bourassa2010predicting} show that spatial structure that is not exploited by39the model surfaces in prediction errors, \citet{pace2020examining} find that tree40ensembles recover information left in the residuals of hedonic and spatial models,41and \citet{roberts2017cross} document that random cross-validation is optimistic42whenever observations are spatially clustered, recommending blocked (spatially43grouped) folds instead. We report both. On the prediction side, \citet{bin2004prediction}44compares parametric and semi-parametric hedonic predictors, \citet{mullainathan2017machine}45frame machine-learning methods as estimators of the conditional expectation46$\mathbb{E}[P\mid\mathbf{z}]$ suited to prediction tasks, and \citet{kok2017bigdata}47document the rise of automated valuation models (AVMs) in industry. This literature48generally finds that flexible learners beat linear hedonics on accuracy at some cost49in interpretability; our estimator ladder (Section~\ref{sec:benchmark}) reproduces50that pattern on a common feature set, and our focus is on how much of the gap is51closed not by a better \emph{estimator} but by better \emph{measurement}.5253The third strand is the use of \textbf{unstructured listing content}---text and54images---in real-estate valuation. \citet{gentzkow2019text} survey text-as-data55methods in economics, and \citet{dell2025deep} and \citet{korinek2023generative}56survey, respectively, deep learning and generative models as research tools for57economists. In real estate specifically, \citet{nowak2017textual} show that words in58MLS remarks predict transaction prices and reduce hedonic pricing errors;59\citet{nowak2021dictionaries} show that the price-relevant vocabulary is not stable60across markets or over time, a warning against transferring word lists; and61\citet{shen2021information} extract quality signals from property descriptions with62supervised machine learning, finding that the text carries information about63unobserved quality that is capitalized into prices. \citet{zhang2024describe}64benchmark text encoders for price prediction and find that textual and65non-textual features are complementary, and a companion working paper on Quebec66listings \citep{boucher2026decoding} documents that semantic embeddings of67descriptions carry out-of-sample price information and that the language of the68description is a first-order confounder in a bilingual market---a finding we69build on here. On the image side, \citet{ahmed2016house} combine visual and70textual features for price estimation; \citet{bessinger2016quantifying} and71\citet{poursaeed2018vision} show that learned measures of exterior and interior72quality carry independent predictive power; \citet{naik2017computer} and73\citet{glaeser2018vision} use street-level imagery to measure the physical74appearance of neighbourhoods and houses and study how appearance relates to75incentives and change; \citet{law2019take} and \citet{kang2021understanding}76combine street-view and satellite imagery with structured attributes to predict77prices and appreciation; \citet{zhang2018greenery} price street-visible greenery78from street-view images; \citet{johnson2020curb} construct a curb-appeal index from79Google Street View images and estimate its capitalization into sale prices; and80\citet{lindenthal2021architecture} classify architectural styles from photographs81and document style premia. Relative to this work, our approach differs in the82\emph{technology of measurement}: rather than training bespoke supervised models on83labelled data---which requires labels and task-specific engineering, and yields84opaque scores---we use general-purpose large language and vision--language models85with schema-constrained decoding to produce a fixed set of interpretable,86human-readable attributes (``renovated kitchen,'' ``brick façade,'' a 1--1087curb-appeal score) at a marginal cost of a fraction of a cent per listing, and we88embed those attributes in a transparent hedonic regression whose coefficients can be89read as implicit prices. The reliability of such zero-shot annotation is itself an90active research question; \citet{gilardi2023chatgpt} find that general-purpose LLMs91match or exceed crowd workers on standard text-annotation tasks, but the extraction92error of any particular field must be assessed, which we do by confronting the93machine-generated attributes with independent sources94(Section~\ref{subsec:validation}).9596The fourth strand is the \textbf{microstructure of listing markets}, which frames two97institutional features of our data. Our sample pools a brokered MLS platform98(Realtor.ca) with a for-sale-by-owner platform (DuProprio); \citet{hendel2009mls}99compare exactly these two market forms and document differences in seller100composition, marketing, and outcomes, motivating our platform indicator and101platform-specific robustness checks. And because we observe asking rather than102transaction prices, the interpretation of our dependent variable rests on the103list-price literature: the asking price is a strategic choice that anchors search and104bargaining \citep{horowitz1992list,knight2002listing,han2015microstructure}, and105\citet{han2016asking} show that it acts simultaneously as a commitment device and106as a signal in a directed-search environment. We return to this limitation in107Section~\ref{subsec:limits}.108