Python 52.4%
TeX 47.6%
1% ============================================================================2\section{Robustness and Limitations}3\label{sec:robustness}4% ============================================================================56\subsection{Robustness}7\label{subsec:robust}89Several features of the design guard against spurious results. The dual10specification (Section~\ref{subsec:econometrics}) ensures that estimates are not11driven by the non-random availability of living area: coefficients on attributes12common to both models retain their sign and rough magnitude whether or not area is13conditioned on. The ``\texttt{has\_}'' indicators accompanying each imputed field let14the regression separate observed from imputed information, so median imputation does15not silently attenuate the size coefficients. The price-per-square-foot trim removes16the residual parsing errors that would otherwise dominate the tails, and the17municipality fixed effects absorb the first-order spatial confound emphasized by18\citet{can1992spatial} and \citet{anselin1988spatial}.1920\paragraph{Inference.} Because listings within a municipality share unobserved21local shocks, our preferred standard errors are clustered at the municipality level22(195 clusters; \citealp{cameron2015cluster}), the same level as the fixed effects.23Clustering matters exactly where theory predicts: the text- and vision-derived24quality attributes---which vary house by house within a municipality---are25essentially unaffected (curb appeal, renovation level, luxury, garage, materials,26landscaping, and the rural-setting discount all remain significant at the 5\% level27or better in Model~A, as do the LLM condition scale, renovated kitchen, and hardwood28indicators), whereas keyword amenities with a strong spatial component lose29precision (public-transit proximity falls to the 5\% boundary in Model~A and is30insignificant in Model~B; central air conditioning is insignificant in Model~B; the31curb-appeal score, precisely estimated in Model~A, loses significance in the broader32Model~B).33Heteroskedasticity-robust (HC3) standard errors34\citep{white1980heteroskedasticity,mackinnon1985some} yield the same qualitative35pattern with uniformly smaller intervals. This asymmetry is reassuring for our main36claim: the AI-generated attributes capture within-market, property-level quality37variation rather than repackaged location effects.3839Table~\ref{tab:robustness} probes the stability of the predictive blend across40subsamples. Three patterns stand out. First, accuracy is comparable whether living41area is observed (median error 13.5\%) or imputed (14.4\%), confirming that the42area-free model prices the incomplete listings almost as well as the rich model.43Second, the model fits the broker-listed Realtor.ca inventory better44($R^2_{\text{OOS}}=0.80$) than the smaller, noisier for-sale-by-owner DuProprio45segment (0.58), though median errors are identical at 14.0\%; the gap is consistent46with the compositional and marketing differences between brokered and FSBO47platforms documented by \citet{hendel2009mls}. Third, the median error is stable48across price terciles (11--16\%), lowest in the mid-market and modestly higher in49the tails, where idiosyncratic and luxury attributes are harder to capture. (Within50a single tercile $R^2$ is uninformative because the price range is restricted; the51median error is the comparable metric.)5253\input{tables/robustness}5455\subsection{Limitations}56\label{subsec:limits}5758Five limitations qualify the interpretation of our estimates.5960\emph{Asking prices.} The dependent variable is the asking price of properties61listed for sale, not the transaction price. The list price is a strategic choice62that anchors buyer search and bargaining63\citep{horowitz1992list,knight2002listing,han2015microstructure}: it is set with64sale in mind and is highly correlated with the eventual transaction price, but65list-to-sale ratios vary systematically with market conditions and overpricing is66more common in thin segments. Two implications follow. First, our implicit prices67measure how sellers and their agents \emph{price} attributes into the ask, which68converges to buyers' valuation only insofar as the market disciplines the ask; the69economically sensible magnitudes and the monotone validation of70Section~\ref{subsec:validation} suggest the discipline is substantial. Second, any71attribute that shifts the list-to-sale ratio itself (e.g., attributes of hard-to-sell72properties) will carry a coefficient that mixes valuation and listing strategy.73Replicating the analysis on registry transaction prices is the natural next step.7475\emph{Non-causality and content endogeneity.} Descriptions and photographs are76marketing instruments chosen by sellers and agents. Richer descriptions, more77photographs, and higher visual quality correlate with higher-end listing agents and78more valuable properties \citep{nowak2017textual}. The coefficients are therefore79\emph{predictive} associations, not causal effects; they should not be read as the80price change from physically adding a given attribute. We control for two observable81dimensions of marketing effort (photo count and virtual tours), but cannot rule out82that residual marketing quality loads on the AI-derived attributes.8384\emph{Machine-generated features.} LLM and vision outputs can contain errors.85Zero-temperature decoding, strict schemas, and closed enumerations bound this risk86and make it reproducible, and classical measurement error in the extracted scores87would bias their coefficients toward zero---making our estimates conservative---but88we do not perform systematic human validation, and the visual layer judges each89property from a single façade photograph, ignoring interiors.9091\emph{Selection.} Listed properties are not a random sample of the housing stock,92and listings with missing area or room detail are not missing at random (they tend93to be less complete listings). The ``\texttt{has\_}'' indicators and the two-part94design mitigate, but do not eliminate, the second concern.9596\emph{Cross-section.} The data are a single-date snapshot, so the model captures no97market dynamics and the location premia are fixed as of 3~July~2026. Municipalities98with fewer than 20 listings are pooled, reducing spatial resolution where data are99sparse.100