% ==========================================================================
\section{Conclusion}
\label{sec:conclusion}
% ==========================================================================

This paper develops a comprehensive hedonic pricing model for Airbnb
listings across Quebec, Canada, combining OLS estimation, quantile
regression, machine learning, and spatial econometrics. Using 5,000
listings from Inside Airbnb---3,412 with valid prices---we estimate nine
nested specifications, quantile regressions at five percentiles, seven
machine learning models, and three spatial econometric models.

\textit{Key findings.} The full hedonic model achieves $R^2 = 0.274$
with 50 covariates. Bedrooms command a premium of approximately 213\%
per unit, modulated by a concave log-bedrooms effect. The review score
carries a premium of approximately 11\% per 0.1-point increment. Local
listing density within 500m depresses prices ($-0.8$\% per listing),
while density at 1~km shows a positive agglomeration effect.

Quantile regression reveals substantial heterogeneity: the entire-home
premium nearly doubles from $\tau = 0.10$ to $\tau = 0.75$, while
review scores matter more at lower quantiles. The pseudo-$R^2$ declines
from 0.100 to 0.020 across quantiles, confirming that luxury pricing
depends heavily on unobserved factors.

XGBoost achieves the best out-of-sample $R^2$ (0.285, MAPE = 14.3\%),
modestly exceeding the OLS model. The similar performance of parametric
and non-parametric models suggests that the log-linear specification
captures the majority of systematic variation.

\textit{Practical implications.} The price decomposition framework
allows hosts to benchmark their listing against comparable properties.
The spatial analysis highlights that immediate competitive density
depresses prices, suggesting differentiation on amenities and quality
in dense urban clusters. Quantile-specific estimates are useful for
revenue management: luxury-segment hosts should prioritize entire-home
exclusivity and premium location, while budget-segment hosts gain from
structural capacity and review credibility.

\textit{Limitations.} The cross-sectional design limits causal
identification. The price variable reflects the listed price, which may
differ from the transaction price. The original dataset contains only
19 variables, requiring extensive feature engineering from listing names
and amenity counts. Certain optional dependencies (CatBoost, pygam,
SHAP, PySAL) were unavailable, necessitating fallback implementations.
The spatial models rely on a fixed $k$-nearest-neighbours weight matrix
($k = 5$); alternative spatial structures may yield different estimates.

\textit{Future research.} Panel data methods exploiting within-listing
variation would strengthen causal identification. Booking-level
transaction data including cleaning fees and occupancy rates would
permit a more complete analysis. Pre-trained language model embeddings
applied to descriptions and reviews could extract richer text-based
features. Extending the framework to other Canadian provinces would test
external validity.