spb/wp22_uqo

1\section{Limitations and conclusion}\label{sec:conclusion}23\subsection{Limitations}45The data limit what any estimator can do. Characteristics come from a single current roll vintage, so renovations after the sale are invisible and recent new constructions may be excluded; the provider collects no sale below \$50{,}000; the roll's neighbourhood code is unusable, and location is a metric grid rather than an economic neighbourhood; arm's-length status, financing and listing information are absent; and the single date per sale conflates signing and registration. The model is sequential: stage two conditions on stage-one residuals and the intervals omit hedonic-parameter and hyper-parameter uncertainty, which the bootstrap shows to matter for mid-sized municipalities. The hierarchy is administrative; a municipality is shrunk toward its MRC whether or not the MRC is its housing market, and the ablation that removes the MRC level shows that this choice changes municipal monthly changes by a few log points. Seasonality is not modelled. The Monte Carlo design, although rich, is ours, and the two counter-examples were chosen by us. Statistics Canada boundaries could not be obtained, so external comparisons use metropolitan communities.67\subsection{Conclusion}89We set out to measure constant-quality house price movements where the data are thin, and to say how far they can be trusted. The framework integrates known components; the evidence shows that their combination produces municipal monthly indices with errors a fraction of those of medians and cell-level time dummies in simulation, intervals whose coverage is close to nominal when the innovation variance is calibrated across cells, a revision profile that is measured rather than designed away, and aggregates that agree with external indices where the scopes match. It also shows where the approach should not be used: in markets with strong idiosyncratic local dynamics the pooled estimator lags a direct one, and for cells with fewer than fifteen effective observations no estimator reports more than a model-assisted interpolation. The panel, the vintage database and the code are released so that both claims can be checked.10