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