spb/wp14_uqo

1% ============================================================================2\section{Conclusion}3\label{sec:conclusion}4% ============================================================================56We have estimated a hedonic price model for Quebec houses that augments the standard7set of structural, locational, and amenity attributes with two machine-generated8feature layers: structured specifications extracted from free-text descriptions by a9large language model, and visual attributes extracted from façade photographs by a10vision--language model. Schema-constrained decoding renders these outputs as11ordinary, auditable econometric variables, obtained at a marginal cost of a fraction12of a cent per listing. The central message is methodological: general-purpose AI13models, disciplined by strict output schemas, function as a low-cost14feature-engineering technology that measurably expands the characteristics set15available to hedonic analysis---and the discipline must extend to the input, since16the text handed to the model may contain the very price one is trying to explain.1718The gains are economically meaningful but unequal across the two layers. Together19they add 3.8 points of in-sample $R^2$ to an already rich log-linear specification20(0.717 to 0.755 with living area observed), of which 3.0 come from the façade21photograph and 0.8 from the text; they raise the out-of-sample accuracy of every22estimator in our ladder, from OLS to gradient boosting and an embedding-based23residual neural network; and the best blend attains an out-of-sample $R^2$ of 0.77924with a median absolute percentage error of 14.0\% under random folds, 0.734 for25gradient boosting when entire postal areas are held out. The vision layer adds more26to the linear model than switching from OLS to a random forest does. The extracted27attributes carry interpretable implicit prices---curb appeal, renovation level,28luxury cues, and exterior materials all price sensibly---that survive29municipality-clustered inference, a cluster bootstrap, finer postal-code fixed30effects, and a control for the language of the description. The text layer's31contribution, by contrast, is modest once keyword amenities are already in the model,32and part of its measured gain on the for-sale-by-owner platform reflects the asking33price embedded in those descriptions rather than information about the house.3435Three extensions are natural. First, the same schema-constrained approach scales36from a single façade image to a small panel of interior photographs (kitchen,37bathroom, living space), which our design already accommodates; a repeat extraction38of a subsample at a fixed temperature, and a human-labelled validation set, would39also quantify the extraction error that we could only measure indirectly. Second,40the DuProprio coordinates and the postal codes of all listings permit continuous41spatial features---distance to the urban core, to water, or to amenities---that would42refine the geography beyond discrete area effects. Third, and most important for43external validity, linking listings to registry transaction prices would establish44how much of the AI-measured quality premium survives bargaining. For appraisal,45automated valuation, and property-tax assessment---applications that consume hedonic46models at scale---the practical implication is that the unstructured majority of a47listing's information content is no longer out of reach, provided that it is read48with the same care as any other regressor.49