spb/wp23_uqo

1\section{Limitations and conclusion}\label{sec:conclusion}23\subsection{Limitations}45The sample is short: 69 months, of which 24 are history, 31 to 45 forecast origins per horizon and one macroeconomic cycle. Tests have low power, model confidence sets are wide, and the twelve-month results describe one episode. The property-level task values a dwelling whose characteristics are known; it does not forecast which dwellings sell, and selection into transaction is documented rather than corrected. The characteristics come from the assessment roll and omit bedrooms, bathrooms, condition and listing information. The neural architectures are trained with fixed hyper-parameters and few epochs, by design, and their errors vary across seeds by as much as the differences between them. The Task~B target is itself an estimate (the WP22 final index), and the simulation study uses the WP22 simulator, whose assumptions are ours.67\subsection{Conclusion}89Do new forecasting architectures substantially improve residential price forecasts? On this data, no. For the price of a future sale, gradient boosting on hedonic and local market features is the most accurate model at every horizon and the hierarchical hedonic network does not improve on it; for the change of a quality-adjusted market index, the forecastable component at one to six months is the cell's own momentum and seasonality, which a pooled tree model or an autoregression captures, and the spatial graph, the regime gating and the hierarchical intercepts add nothing that survives an ablation; at twelve months nothing beats a simple autoregression and every pooled model inherits the bias of the preceding boom. Two of the proposed architectures earn a qualified place: the quality-adjusted state-space forecaster, because it is stable across seeds and the only model with honest intervals at every horizon on the real data, although it is the worst model in the simulation when the real-time index is noisy; and the integrated architecture, because it is competitive and shows that an equal-weight combination of components is at least as good as a learned stack. The simulation study (Section~\ref{sec:sim}) explains the empirical ranking: where latent dynamics are persistent the pooled models recover the drift, where they are a random walk nothing beats no change, and spatial spillovers are learnable only when they are strong and the sample is long. The value of the exercise is the measured answer and the design that produced it: strict chronology, leakage-audited information sets, inference that respects short samples, and ablations that attribute performance to mechanisms. The code, the panel and every result are released so that the next architecture can be tested the same way.