spb/wp9_uqo

1\begin{table}[htbp]\centering2\caption{Random versus spatial holdouts: valuation accuracy}3\label{tab:spatial_holdout}4\begin{threeparttable}\small5\begin{tabular}{lccc}6\toprule7 & Random 80/20 & \multicolumn{2}{c}{Held-out FSAs (20\%)} \\8\cmidrule(lr){3-4}9Metric & \makecell{(within\\support)} & \makecell{Province-mean\\effect} & \makecell{IDW of 5\\nearest FSAs} \\10\midrule11Out-of-sample $R^2$ (log price) & 0.769 \scriptsize(0.010) & 0.520 \scriptsize(0.034) & 0.743 \scriptsize(0.014) \\12RMSE (log points) & 0.280 \scriptsize(0.005) & 0.402 \scriptsize(0.013) & 0.294 \scriptsize(0.008) \\13Median absolute \% error & 15.5 \scriptsize(0.2) & 25.0 \scriptsize(1.0) & 17.3 \scriptsize(0.6) \\14Mean absolute \% error (top 1\% trimmed) & 20.4 \scriptsize(0.1) & 32.3 \scriptsize(1.0) & 22.5 \scriptsize(0.9) \\15Share within $\pm$10\% & 34.3 \scriptsize(0.4) & 21.2 \scriptsize(1.0) & 31.2 \scriptsize(1.2) \\16Share within $\pm$20\% & 60.3 \scriptsize(0.4) & 40.9 \scriptsize(1.7) & 56.1 \scriptsize(1.6) \\17\bottomrule18\end{tabular}19\begin{tablenotes}\footnotesize\item Notes: Means over five replications, standard deviations across replications in parentheses. Column~1: random 80/20 splits of listings, evaluation restricted to FSAs observed in training. Columns~2--3: 20\% of the FSAs (229 of the non-pooled neighbourhoods) are held out entirely; their listings are valued with the structural coefficients from the training FSAs and a neighbourhood effect imputed either as the mean effect of the training FSAs in the same province or as the inverse-distance-weighted mean of the five nearest training FSAs (centroid distance). Prices back-transformed with Duan's smearing estimator. The trimmed mean excludes the largest 1\% of errors in each replication: a few listings with implausible parking or storey counts (up to 3{,}620 spaces) otherwise dominate the untrimmed mean.20\end{tablenotes}21\end{threeparttable}22\end{table}23