\begin{table}[htbp]
\centering\scriptsize
\caption{Computational cost (wall-clock on a 32-core Mac Studio M3 Ultra; tree models 10 threads, neural models 8 threads)}\label{tab:cost}
\begin{threeparttable}
\begin{tabular}{lrrrrr}
\toprule
Task & Model & Seconds per fit & Total seconds & Fits & Mean training rows \\
\midrule
A & Ridge & 0.1 & 1 & 12 & 314{,}363 \\
A & Hedonic OLS & 0.3 & 4 & 12 & 314{,}363 \\
A & Elastic net & 3.4 & 40 & 12 & 314{,}363 \\
A & Quantile MLP & 7.7 & 93 & 12 & 314{,}363 \\
A & HSHN (full pooling) & 9.0 & 108 & 12 & 314{,}363 \\
A & HSHN (no pooling) & 9.3 & 111 & 12 & 314{,}363 \\
A & HSHN (partial pooling) & 9.5 & 114 & 12 & 314{,}363 \\
A & XGBoost & 9.9 & 119 & 12 & 314{,}363 \\
A & Random forest & 29.9 & 359 & 12 & 314{,}363 \\
A & SVR (subsample) & 63.6 & 764 & 12 & 314{,}363 \\
A & LightGBM & 155.0 & 1860 & 12 & 314{,}363 \\
B & No change & 0.0 & 0 & 36 & 5{,}785 \\
B & AR(2), cell & 0.0 & 0 & 36 & 5{,}785 \\
B & Seasonal naive & 0.0 & 0 & 36 & 5{,}785 \\
B & Mean growth & 0.0 & 0 & 36 & 5{,}785 \\
B & Ridge (panel) & 0.0 & 0 & 36 & 5{,}785 \\
B & Random walk with drift & 0.0 & 0 & 36 & 5{,}785 \\
B & Momentum & 0.0 & 0 & 36 & 5{,}785 \\
B & BHPF & 0.2 & 7 & 36 & 5{,}785 \\
B & QANSS & 0.2 & 8 & 36 & 5{,}785 \\
B & MLP (panel) & 1.0 & 36 & 36 & 5{,}785 \\
B & Local linear trend & 1.6 & 58 & 36 & 5{,}785 \\
B & RAREF & 3.9 & 140 & 36 & 5{,}785 \\
B & DSGFN & 4.9 & 176 & 36 & 5{,}785 \\
B & TFT-lite & 5.0 & 181 & 36 & 5{,}785 \\
B & LightGBM (panel) & 5.8 & 210 & 36 & 5{,}785 \\
B & GRU & 7.1 & 254 & 36 & 5{,}785 \\
B & LSTM & 8.9 & 322 & 36 & 5{,}785 \\
B & QREFA & 10.4 & 373 & 36 & 5{,}785 \\
B & TCN & 186.0 & 6696 & 36 & 5{,}785 \\
\bottomrule
\end{tabular}
\end{threeparttable}
\end{table}