spb/wp4_uqo

1% ==========================================================================2\section{Conclusion}3\label{sec:conclusion}4% ==========================================================================56This paper has estimated hedonic price functions for a province-wide7cross-section of Airbnb listings in Quebec from a scrape of 5,000 listing8cards. Its first finding concerns the data: the displayed price field9mixes nightly rates with the total prices of stays of about three weeks,10for roughly 28\% of priced listings, concentrated among small Montreal11rental units with few reviews. A two-regime mixture of hedonic12regressions separates the two populations sharply and leaves 2,48013listings with nightly prices. Every estimate in the first version of this14paper was computed on the mixture; none survives, and the magnitudes15reported there (a bedroom premium above 200\%, an 11\% premium per tenth16of a rating point, negative ``amenity index'' effects) were artefacts of17the contamination and of regressors that should not have been18constructed. We have documented these corrections in full so that the19lesson is usable by others who work with scraped platform prices: the20bimodality was visible in the first summary table, and a ratio of21medians by bedroom count is enough to diagnose it.2223On the nightly-price sample the hedonic model behaves as the literature24leads one to expect. Capacity, bathrooms and bedrooms, the shared-bathroom25indicator that stands in for the unobserved room type, the number of26photos and the displayed rating are the priced attributes; title27categories matter once location is held fixed, with chalets and houses28priced about 8 to 10\% above rental units of the same capacity; local29density is positively priced within cities. The Superhost and Guest30Favourite badges carry no premium conditional on capacity, type and31location, and the number of reviews is unrelated to price on average.32The specification with city fixed effects explains half of the variance33of log prices, with a root mean squared error of 0.475 log points, and34its coefficients are stable to the choice of fixed-effect unit, to35clustering, to the regime cut-off and to the exclusion of commercial36lodging or of Montreal.3738The three complementary estimators each add something the linear model39does not show. Quantile regression reveals that cheap listings are priced40per guest and certified by reviews and the Superhost badge, whereas41expensive listings are priced per room and per bathroom, closer to the42centre, and are cheaper when heavily reviewed; the inter-quantile43differences for bedrooms, bathrooms and review counts are statistically44significant. The spatial models find moderate spatial dependence45($\hat\rho \approx 0.20$--$0.26$, $\hat\lambda \approx 0.28$) and a spatial46multiplier of about 1.25, but the direct impacts are within a few47percent of the OLS coefficients, so the hedonic estimates are not48materially biased by the omitted spatial lag. The machine-learning49benchmarks improve the out-of-sample $R^2$ from 0.47 to 0.56 under random50cross-validation and from 0.41 to 0.48 when whole areas are held out; the51seven-point loss under spatial blocking applies to every model, and the52SHAP decomposition of the best learner ranks capacity first and the53badges last, as the hedonic coefficients do, and attributes to the54review count a weight that the linear model cannot, because its55association with price changes sign along the distribution. The flexible models refine the hedonic surface56rather than contradict it, and a median pricing error of about one57quarter of the nightly price is the practical limit of valuation from58card-level information.5960\textit{Limitations.} The price is an asking price displayed on a single61day and excludes the cleaning fee; the data are a cross-section, so62seasonality and within-listing dynamics are unobserved and all63coefficients are conditional associations rather than marginal64willingness to pay. The room type, the host identifier and the amenity65list were not scraped, so professionalisation and specific amenities66cannot be priced, and the stay-regime listings cannot be converted to67nightly rates. The scrape reflects the platform's search results on the68collection dates and is denser in urban and resort markets.6970\textit{Future work.} A repeated scrape that records the search dates,71the stay length behind each displayed price, the room type and the host72identifier would remove the main measurement problems and allow panel73estimation of within-listing price responses to reviews, badges and74seasons. Linking listings to realised bookings or to the calendar would75replace asking prices with revenue. Finally, the Quebec market offers a76natural setting to study how Montreal's registration and77primary-residence rules shift the composition of listings and their78prices, a question that the cross-sectional stay-price pattern documented79here already hints at.80