spb/wp2_uqo

1% Author: Simon-Pierre Boucher — contact@spboucher.ai2%3%======================================================================4\section{Discussion} \label{sec:discussion}5%======================================================================67\subsection{What the Evidence Establishes}89Three findings survive every specification in this paper. First, broker-written descriptions contain price-relevant information beyond the structural fields. The 20 semantic similarities are jointly significant in the baseline, within language subsamples, within forward sortation areas and municipalities, with clustered standard errors, after correcting the lot-size measure, on the floor-area subsample, under every perturbation of the reference texts and with both encoders; and the information generalises out of sample, to random held-out listings and to held-out areas alike. Second, the magnitude of this information is smaller than the first version of the paper suggested and depends on what one controls for. The headline gain of six percentage points of $R^2$ over a structural model with no location variables shrinks to about two points once the language of the listing and area fixed effects are accounted for with the baseline encoder, and to about four points with a multilingual encoder; location explains far more of the variation in asking prices than text does. Third, the individual implicit price gradients are robust for a core of five dimensions---Luxury, Land \& Nature, Family-Friendly, Quiet \& Peaceful, Needs Renovation---and not for the others.1011The interpretable dimensions tell a coherent story. Listings that use luxury vocabulary are priced higher conditional on structure, in every specification and with both encoders, with a gradient of roughly 4 to 14\% per standard deviation depending on the controls; the quantile regressions show the gradient rising with the price level, consistent with \citet{zietz2007determinants} and \citet{mcmillen2008changes}. Listings that describe a natural setting are priced higher; the gradient is large in the baseline (+13\%) and falls to +4\% once the lot area is measured correctly, which indicates that the ``large wooded lot'' reference was in part a proxy for lot size and in part a genuine amenity signal---a decomposition that structured data alone could not have provided. Using Quebec City transactions and site inspections, \citet{desrosiers2002landscaping} estimate landscaping premiums of roughly 4 to 12\%; the text-based gradient, net of lot area, lies at the lower end of that range. Listings that need renovation are priced lower, with a gradient of 4 to 8\%, the sign expected from the capitalisation of deferred maintenance \citep{nowak2017quality, goodwin2014broker}. Listings that market a family-friendly or quiet setting are priced lower, in every specification including those with area fixed effects, which means that these words mark lower-priced segments within areas as well as between them; the equilibrium-sorting logic of \citet{kuminoff2013new} and the evidence of \citet{delmelle2021language} that advertisement text predicts neighbourhood composition both point the same way: listing language is as much about where a property is as about what it is.1213\subsection{What the Evidence Does Not Establish}1415The first version of this paper reported an 8.5\% urgency discount for Motivated Seller language and compared it with the 3.7\% premium that \citet{levitt2008information} find for agent-owned homes. The comparison does not survive the revision. Motivated Seller is the dimension with the largest French--English similarity gap, and its coefficient falls to an insignificant $-2.9\%$ once language is controlled, to $-4.0\%$ (marginally significant) within FSAs, and to zero with the multilingual encoder. The listing-language literature does find that remarks signalling seller motivation are associated with lower prices and faster sales \citep{haag2000real, goodwin2014broker}; our data do not allow us to confirm a robust magnitude for Quebec asking prices. Similarly, the counter-intuitive signs on New Construction ($-10.5\%$) and Entry-Level (+10.4\%) in the baseline reverse with the multilingual encoder (+7.2\% and $-4.9\%$), and the negative signs on Bright \& Spacious and Energy Efficient reverse as well. These reversals are not evidence that one encoder is right and the other wrong; they are evidence that, in a block of 20 projections correlated at 0.66--0.97, the coefficient of any one dimension is defined by the vocabulary that happens not to be shared with the other nineteen, and that this residual vocabulary differs across encoders and across phrasings. The reference-phrasing experiment of Section~\ref{sec:extended} makes the same point from a different direction.1617The design itself contributes to the problem. Appendix~\ref{app:references} shows that several references share phrases---Land \& Nature includes ``bord de l'eau, acc\`es au lac'' (Waterfront), Modern/Contemporary includes ``construction neuve'' and ``\'eco\'energ\'etique'' (New Construction, Energy Efficient), Premium Location includes ``pr\`es des \'ecoles, des parcs, quartier familial'' (Family-Friendly). The principle of dimensional specificity stated in Section~\ref{sec:methodology} was not fully respected, and these overlaps are a plausible mechanism for the sign anomalies: when Modern/Contemporary absorbs the positive valuation of new and efficient construction, New Construction and Energy Efficient are left with the residual association of peripheral new developments. A redesign with disjoint vocabularies, or an explicit orthogonalisation of the references in embedding space, would be the first step for future applications.1819\subsection{Language as a Confound in Multilingual Text-as-Data Applications}2021The most consequential finding of the revision is methodological. Reference-based projection, like any dictionary method, compares each document with a fixed anchor; when the corpus is bilingual and the anchors are monolingual, the comparison measures language before it measures content. In our data this single confound accounts for 86\% of the variance of the mean similarity and for more than half of the apparent semantic gain, and it is correlated with price through the geography of the two language communities. The problem is not specific to housing: it arises whenever a text-as-data study pools documents written in different languages, or in different registers, and scores them against a single set of references \citep{gentzkow2019text, ash2023text}. Three remedies are available and we have applied all three: control for language explicitly, estimate within language, and use an encoder trained to align languages \citep{reimers2020multilingual}. The multilingual encoder is the most satisfactory because it removes the gap at the source and retains more information, but it also changes the individual coefficients, which shows that the choice of encoder is part of the specification and should be reported and varied as such.2223\subsection{Relation to Prior Findings}2425On the information content of text, our results agree with \citet{nowak2017quality}, who show that words in agent comments proxy for unobserved quality, and with \citet{shen2020text}, who find substantial information about unobserved quality in property descriptions; our out-of-sample results quantify this information for Quebec asking prices at about four percentage points of $R^2$ and a fraction of a point of median absolute percentage error, net of language. On the magnitude of text gains relative to unrestricted representations, our principal-component and multilingual benchmarks reproduce the ordering reported by the prediction-oriented literature \citep{shen2020text, baur2023automated}: unrestricted features fit better, and the reference approach pays a fit penalty of about two points for its names. On the stability of text features across contexts, \citet{nowak2021dictionaries} show that real estate dictionaries differ across space and time; our finding that the same references yield different coefficients across languages and encoders is the embedding-space analogue of their result. On distributional heterogeneity, the rising Luxury gradient agrees with \citet{zietz2007determinants} and \citet{mcmillen2008changes}. On the parallel with listing photographs \citep{glaeser2018computer, poursaeed2018vision}, both modalities reveal quality that structured fields omit, and both face the attribution problem we document when their features are many and correlated.2627\subsection{Description Length}2829The positive coefficient on description length (+10.7\% in the baseline, +7.9\% within FSAs, +6.8\% with the corrected lot area) is one of the more stable text effects in the paper. Conditional on what a description says and on the language it is written in, a longer description is associated with a higher asking price, which is consistent with length proxying for broker effort, property complexity or an abundance of features worth describing \citep{shen2020text, goodwin2014broker}. The variable is truncated at approximately 700 characters, so the coefficient measures length variation within that range.3031\subsection{Practical Implications}3233For automated valuation, the results are encouraging but measured. Text features computed with a multilingual encoder improve out-of-sample $R^2$ by about three percentage points over a model with structure, length and language, and by about seven points over a structural model, at negligible computational cost; but the dominant sources of valuation error in these data are the absence of floor area and of fine location, and an AVM should address those first. Reference-based features have the advantage over principal components that they can be audited---an appraiser can read the reference that defines each one---but the disadvantage, documented here, that their individual weights are not stable to the phrasing of the references. For researchers, the recommended practice that emerges is to (i) report the language composition of the corpus and control for it or use a language-aligned encoder; (ii) design references with disjoint vocabularies; (iii) report block- and domain-level tests as the primary inferential objects; (iv) report the sensitivity of individual coefficients to reference phrasing; and (v) validate out of sample with folds that respect the spatial structure of the data \citep{roberts2017cross}. For brokers, the correlational nature of the estimates bears repeating: the coefficients describe how asking prices and descriptions co-vary, not what rewriting a description would do to the price a property fetches.3435\subsection{Limitations}3637\textbf{Asking prices.} The data are asking prices from a single snapshot of active listings. List and sale prices differ systematically, and the gap varies with market conditions and marketing time \citep{haurin2010list}; if the gap varies with the semantic content of the description---plausibly larger for listings that disclose urgency---the coefficients on sale prices would differ from those reported here. The estimates describe the pricing behaviour of sellers and brokers conditional on the text they wrote.3839\textbf{Identification.} The cross-sectional design cannot separate the information, persuasion and omitted-location channels of Section~\ref{sec:identification}, and the fixed effects absorb only between-area location. Text is written after the asking price is set and by the same agent who sets it, so the estimates are associations in a jointly determined pair of outcomes. Within-property variation across successive listings, or platform experiments on listing text, would be required for causal statements; \citet{alfano2022word} point toward designs that exploit variation in listing composition.4041\textbf{Missing structural information.} The baseline lacks floor area and measures lot size with mixed units. The revision corrects the second and addresses the first on a 36\% subsample; on that subsample the semantic block's increment is about one percentage point of $R^2$, and a full-sample estimate with floor area would likely be of that order.4243\textbf{Encoder and language.} The baseline encoder was trained mainly on English and the references are French; 40\% of the corpus is English. We have documented the consequences and provided multilingual estimates, but a French-specific or domain-adapted model \citep{martin2020camembert, le2020flaubert} might extract more; \citet{conneau2020unsupervised} document material performance losses when general multilingual models are applied to specialised monolingual tasks.4445\textbf{Reference design.} The 20 references are one operationalisation, with overlapping vocabularies, and the coefficients attached to them are sensitive to phrasing. The paper quantifies this dependence but does not resolve it; a principled procedure for designing or orthogonalising references is the natural next step.4647\textbf{Collinearity.} The 20 similarities are correlated at 0.66--0.97 with variance inflation factors averaging 17. Block- and domain-level conclusions are unaffected; dimension-level conclusions are, and we have restricted our claims accordingly.4849\textbf{Temporal stability.} The data are a single cross-section from January 2026. The relation between text and prices may vary over the cycle---urgency language, for instance, may be more diagnostic when it is rarer---and longitudinal data would be needed to assess this.50