spb/wp23_uqo

1\begin{table}[htbp]2\centering\footnotesize3\caption{Hyper-parameters}\label{tab:hyper}4\begin{threeparttable}5\begin{tabular}{p{3.4cm}p{11.5cm}}6\toprule7Model & Hyper-parameters (fixed in advance, never tuned on test origins) \\8\midrule9HSHN & embeddings 8 (geo) / 4 (h); encoder 256-128 GELU, dropout 0.1; AdamW, one-cycle lr $2\times10^{-3}$, 8 epochs, batch 4,096; municipality weight decay $5\times10^{-3}$ \\10MLP (A) & 256-128 GELU, dropout 0.1; pinball loss on 5 quantiles; same optimiser \\11LightGBM (A) & 800 trees, lr 0.04, 63 leaves, min child 30, subsample 0.8; quantile models 400 trees \\12XGBoost (A) & 600 trees, lr 0.05, depth 8, subsample 0.8 \\13Random forest (A) & 150 trees, depth 18, min leaf 5, max features 0.5 \\14SVR (A) & RBF, C=3, $\epsilon$=0.05, 15,000-sale subsample \\15Panel MLP / LSTM / GRU / TCN / TFT-lite (B) & hidden 64 (MLP) / 32 (RNN, attention) / 16 channels (TCN); 60 epochs, batch 512, AdamW lr $2\times10^{-3}$, weight decay $10^{-4}$; pinball loss \\16DSGFN (B) & node encoder 32; one attention layer; prior adjacency: Gaussian distance kernel (60 km) $\times$ co-movement; 60 epochs \\17RAREF (B) & 3-state Gaussian HMM (60 EM iterations) on provincial real-time growth; 3 ridge experts ($\alpha$=3) \\18BHPF (B) & conjugate Gaussian hierarchy; 15 variance-component iterations; priors $\tau\_b^2$=1 (slopes) \\19QANSS (B) & 32 tanh units; 200 Adam steps, lr $5\times10^{-3}$; Gaussian likelihood \\20QREFA (B) & NNLS stacking of QANSS, BHPF, DSGFN (40 epochs), RAREF on training origins \\21LightGBM (B) & 300 trees, lr 0.03, 15 leaves, min child 40; quantile models \\22Ridge (B) & $\alpha$=3 on standardised features \\23\bottomrule24\end{tabular}25\end{threeparttable}26\end{table}27