spb/wp14_uqo

1% Auto-generated (revision: full ladder in both panels)2\begin{table}[htbp]3  \centering4  \begin{threeparttable}5    \caption{Predictive Performance: From OLS to Deep Learning}6    \label{tab:models}7    \begin{tabular}{@{}lcccc@{}}8      \toprule9      Estimator & $R^2_{\text{OOS}}$ & Median AE & Within $\pm$15\% & Within $\pm$25\% \\10      \midrule11      \multicolumn{5}{@{}l}{\textit{Panel A: Structural + amenity + LLM-text features}} \\[3pt]12      OLS (linear hedonic)              & 0.628 & 17.9\% & 43\% & 64\% \\13      Random Forest                     & 0.677 & 17.9\% & 43\% & 64\% \\14      Neural network (plain MLP)        & 0.711 & 16.1\% & 47\% & 69\% \\15      Neural network (robust ResNet)    & 0.741 & 15.2\% & 49\% & 70\% \\16      Gradient boosting (HistGBM)       & 0.743 & 15.4\% & 49\% & 70\% \\17      Blend (NN + GBM)                  & 0.760 & 14.7\% & 51\% & 72\% \\[4pt]18      \multicolumn{5}{@{}l}{\textit{Panel B: \ + Vision (façade) features}} \\[3pt]19      OLS (linear hedonic)              & 0.686 & 16.9\% & 45\% & 67\% \\20      Random Forest                     & 0.704 & 17.2\% & 45\% & 66\% \\21      Neural network (robust ResNet)    & 0.747 & 14.8\% & 51\% & 71\% \\22      Gradient boosting (HistGBM)       & 0.770 & 14.5\% & 51\% & 71\% \\23      \textbf{Blend (NN + GBM)}         & \textbf{0.779} & \textbf{14.0\%} & \textbf{53\%} & \textbf{73\%} \\24      \bottomrule25    \end{tabular}26    \begin{tablenotes}[flushleft]27      \footnotesize28      \item \textit{Notes:} Out-of-sample performance from 5-fold cross-validation with a29      common seed (identical folds across rows). $R^2_{\text{OOS}}$ is on30      $\ln(\text{Price})$; ``Median AE'' is the median absolute percentage error in31      levels (Equation~\ref{eq:cvmetrics}). Panel~A uses structural, amenity, room-detail,32      listing-signal, and LLM text-derived features; Panel~B adds the façade33      vision layer (21 regressors derived from the twelve schema fields). City, type, and source enter as one-hot indicators for the34      linear and tree models and as learned embeddings for the neural networks. The tree35      and neural models handle missing living area natively; OLS uses within-type median36      imputation. The blend weights the two models equally (weight fixed a priori). The best37      model is the text\,+\,vision blend. Entries may differ by $\pm0.001$ from38      Tables~\ref{tab:ablation} and~\ref{tab:rob_spatial}, which come from separate runs.39    \end{tablenotes}40  \end{threeparttable}41\end{table}42