\begin{table}[htbp]
\centering\footnotesize
\caption{Hyper-parameters}\label{tab:hyper}
\begin{threeparttable}
\begin{tabular}{p{3.4cm}p{11.5cm}}
\toprule
Model & Hyper-parameters (fixed in advance, never tuned on test origins) \\
\midrule
HSHN & 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 $5\times10^{-3}$ \\
MLP (A) & 256-128 GELU, dropout 0.1; pinball loss on 5 quantiles; same optimiser \\
LightGBM (A) & 800 trees, lr 0.04, 63 leaves, min child 30, subsample 0.8; quantile models 400 trees \\
XGBoost (A) & 600 trees, lr 0.05, depth 8, subsample 0.8 \\
Random forest (A) & 150 trees, depth 18, min leaf 5, max features 0.5 \\
SVR (A) & RBF, C=3, $\epsilon$=0.05, 15,000-sale subsample \\
Panel 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 \\
DSGFN (B) & node encoder 32; one attention layer; prior adjacency: Gaussian distance kernel (60 km) $\times$ co-movement; 60 epochs \\
RAREF (B) & 3-state Gaussian HMM (60 EM iterations) on provincial real-time growth; 3 ridge experts ($\alpha$=3) \\
BHPF (B) & conjugate Gaussian hierarchy; 15 variance-component iterations; priors $\tau\_b^2$=1 (slopes) \\
QANSS (B) & 32 tanh units; 200 Adam steps, lr $5\times10^{-3}$; Gaussian likelihood \\
QREFA (B) & NNLS stacking of QANSS, BHPF, DSGFN (40 epochs), RAREF on training origins \\
LightGBM (B) & 300 trees, lr 0.03, 15 leaves, min child 40; quantile models \\
Ridge (B) & $\alpha$=3 on standardised features \\
\bottomrule
\end{tabular}
\end{threeparttable}
\end{table}