spb/wp23_uqo

1\section{Ablations and robustness}\label{sec:ablation}23\subsection{Component ablations of the proposed architectures}45Table~\ref{tab:ablation_b} removes one component at a time from the Task~B architectures at the municipal level and reports the change in MAE with a stationary-bootstrap interval over the per-origin loss differences. The results are the most informative of the paper. For DSGFN, removing the prior graph entirely (uniform prior, attention still learned) does \emph{not} hurt: the variant is as accurate at one and three months and better at six months (\AblDiffdsgfnnographhSix{} log points, interval excluding zero), and removing the distance kernel or the co-movement term changes little. The neighbourhood vector that DSGFN learns therefore carries no information beyond what the parent-momentum features already give every pooled model: on 64 municipal nodes and 45 origins, the attention learns a weighted average of other cells' histories that is at best a noisier version of the regional and provincial momentum. The attention diagnostics confirm the reading: the top neighbour receives a weight of \AttTopThree{} at three months, the attention entropy is close to that of a uniform distribution over the class's cells, the top neighbour is in the same region only \AttSameRegThree\% of the time and in the same MRC \AttSameMrcThree\% of the time, with a median distance of \AttDistThree{} km. For QANSS, removing the parent drifts from the state changes the error by \AblDiffqanssnoparenthThree{} log points at three months, within the bootstrap interval: the learned transition uses the cell's own recent drift. For the integrated architecture, the equal-weight combination of the four components beats the non-negative least-squares stack at every horizon (\AblDiffequalweightcombohThree{} log points at three months, \AblDiffequalweightcombohSix{} at six), the classical forecast-combination result \citep{batesgranger1969combination,timmermann2006combinations}: with 45 origins, the stacking weights overfit. Removing any single component from the stack changes the error by less than the bootstrap interval, except BHPF at three and six months.67The feature-group ablations of the pooled ridge model locate the predictable signal: removing the cell's own momentum features raises the error by \AblDiffridgenomomentumhThree{} log points at three months (interval excluding zero) and more at six months, removing the volume features changes it by \AblDiffridgenovolumehThree{}, and removing the parent (regional and provincial) features by \AblDiffridgenoparenthThree{}. Own momentum is the forecastable component; the parent momentum adds a little at short horizons and nothing at six months once the own momentum is present.89\input{tables/tab_ablation_b}1011\subsection{Robustness}1213Table~\ref{tab:robust_b} re-runs a subset of Task~B models at the three-month horizon under alternative designs. Four facts are stable. Rolling 24-origin windows help the pooled ridge (\RobridgepanelrollingTwoFour{} against \Robridgepanelexpanding{} log points, as parameter drift in a short sample favours recent data) and hurt the neural models, which need the observations. Dropping the thinnest cell-months ($n_{\text{last}} < 10$) or winsorising the targets changes every model by less than 0.2 log points and leaves the ranking intact. The DSGFN bandwidth (30 or 120 km instead of 60) is irrelevant, consistent with the ablation that the graph carries no information. Most importantly, re-estimating the neural models with different seeds moves their error by as much as the differences between models: DSGFN ranges from \RobdsgfnseedTwo{} to \RobdsgfnseedOne{} log points across three seeds and TFT-lite from \Robtftliteexpanding{} to \RobtftliteseedTwo, while the ridge and AR benchmarks have no seed. With 45 origins and 64 cells the neural models are not reliably distinguishable from one another, and any ranking among them in Section~\ref{sec:results} should be read with this seed variability in mind; the proposed QANSS is the most stable of the three (\Robqanssexpanding{}, \RobqanssseedOne{}, \RobqanssseedTwo).1415\input{tables/tab_robust_b}16