\begin{table}[htbp]
\centering\scriptsize
\caption{Simulation — MAE of 3-month index-change forecasts relative to the no-change forecast at the municipality level, by scenario (Monte Carlo s.e. over replications)}\label{tab:sim_municipality}
\begin{threeparttable}
\begin{tabular}{lrrrrr}
\toprule
Model & rw & momentum & regimes & spillover & sparse \\
\midrule
No change & 1.00 (0.00) & 1.00 (0.00) & 1.00 (0.00) & 1.00 (0.00) & 1.00 (0.00) \\
Mean growth & 0.78 (0.05) & 1.13 (0.05) & 0.82 (0.09) & 0.98 (0.05) & 1.13 (0.05) \\
Momentum & 0.98 (0.01) & 1.00 (0.01) & 0.87 (0.03) & 0.93 (0.01) & 1.00 (0.01) \\
Random walk with drift & 1.11 (0.06) & 1.18 (0.06) & 0.80 (0.06) & 0.97 (0.03) & 1.26 (0.07) \\
Seasonal naive & 2.21 (0.16) & 2.04 (0.12) & 1.36 (0.13) & 1.51 (0.06) & 2.36 (0.18) \\
AR(2), cell & 0.76 (0.05) & 1.07 (0.05) & 0.78 (0.08) & 0.91 (0.03) & 1.07 (0.05) \\
Local linear trend & 2.29 (0.15) & 1.90 (0.11) & 1.13 (0.08) & 1.33 (0.04) & 2.26 (0.17) \\
Ridge (panel) & 1.13 (0.13) & 1.53 (0.08) & 1.20 (0.20) & 1.21 (0.10) & 1.58 (0.12) \\
LightGBM (panel) & 0.87 (0.06) & 1.16 (0.05) & 0.79 (0.07) & 0.99 (0.04) & 1.16 (0.06) \\
MLP (panel) & 0.91 (0.08) & 1.29 (0.04) & 0.89 (0.11) & 1.17 (0.13) & 1.36 (0.14) \\
LSTM & 0.90 (0.11) & 1.18 (0.05) & 0.91 (0.12) & 1.08 (0.11) & 1.33 (0.13) \\
TCN & 0.88 (0.11) & 1.21 (0.06) & 0.91 (0.12) & 1.07 (0.12) & 1.31 (0.13) \\
TFT-lite & 0.83 (0.09) & 1.18 (0.05) & 0.89 (0.16) & 0.99 (0.09) & 1.19 (0.10) \\
BHPF & 1.16 (0.14) & 1.61 (0.08) & 1.22 (0.19) & 1.22 (0.11) & 1.65 (0.13) \\
RAREF & 1.19 (0.14) & 1.54 (0.08) & 1.25 (0.19) & 1.23 (0.14) & 1.60 (0.09) \\
QANSS & 2.04 (0.23) & 1.51 (0.09) & 1.30 (0.25) & 1.15 (0.06) & 1.57 (0.09) \\
DSGFN & 1.08 (0.10) & 1.32 (0.08) & 0.95 (0.12) & 1.29 (0.13) & 1.37 (0.13) \\
\bottomrule
\end{tabular}
\begin{tablenotes}\footnotesize\item rw: random walk with a constant drift; momentum: AR(1) drifts at every level; regimes: Markov-switching provincial drift; spillover: municipal drift follows the lagged regional drift; sparse: momentum with 35\% of the sales. Values below 1 beat the no-change forecast. Classic and tree models: 10 replications; neural models: 6.\end{tablenotes}
\end{threeparttable}
\end{table}