spb/wp23_uqo

1\section{Introduction}\label{sec:intro}23Forecasting residential prices is a different problem from measuring them. A price index answers how much the price of a constant-quality dwelling has changed; a forecast answers how much it will change, or what a given dwelling will sell for, using only what is known at the forecast origin. The two problems share the data and the quality-adjustment apparatus, but they differ in what can go wrong: an index can be revised, a forecast cannot be, and an index is judged against a latent truth while a forecast is judged against an outcome that arrives later. This paper takes the transaction database, the geography and the real-time index vintages constructed in UQO Working Paper 22 \citep{boucher2026sharphpi} and asks whether forecasting architectures that combine property-level characteristics, spatial dependence, hierarchical pooling, regime adaptation and probabilistic output improve on conventional forecasts, and when.45The recent literature offers many candidate mechanisms and few controlled comparisons. Gradient boosting dominates cross-sectional valuation contests \citep{mayer2019updating,hong2020randomforest}, but valuation is not forecasting; cross-regional spatial dependence of price dynamics is well documented \citep{pace2000spatiotemporal,hollypesaran2010,brady2011diffusion}, and graph neural networks designed for traffic forecasting \citep{li2018dcrnn,wu2019graphwavenet,bai2020agcrn} offer a flexible way to learn it, but the housing applications so far value properties cross-sectionally \citep{zhang2021mugrep,das2021gsne}; regime switching in house price dynamics is classical \citep{hamilton1989regime,capozza2004anatomy}; hierarchical Bayesian pooling of local indices is established \citep{francke2017hierarchical}; neural state-space models exist in the forecasting literature \citep{rangapuram2018dssm}; and simple linear models have been shown to beat elaborate transformers on many benchmarks \citep{zeng2023transformers}. What is missing is an evaluation of these mechanisms on the same transaction-level data, in a design that uses only information available at each origin, with tests that account for overlapping horizons and serial dependence, and with a simulation study that explains why a mechanism helps when it does.67We make four contributions. First, we define three forecasting tasks on an audited registry sample and construct, for every monthly origin from December 2022 to September 2025, the information set actually available then: property characteristics, real-time municipal market features computed from past sales, and the real-time quality-adjusted index vintages of WP22, which were re-estimated at every publication date so that no full-sample index enters a predictor. Second, we implement five candidate architectures and an integrated architecture alongside \NBenchB{} benchmarks, from no-change and momentum rules through ARIMA-type, factor and state-space models to gradient boosting and recurrent, convolutional and attention networks, each with a quantile or distributional output. Third, we evaluate them with forward-chaining experiments and the inferential toolkit of the forecasting literature: Diebold--Mariano tests with horizon-adjusted long-run variances \citep{dieboldmariano1995,harvey1997equality}, Giacomini--White conditional tests \citep{giacominiwhite2006}, model confidence sets \citep{hansen2011mcs}, stationary block bootstraps \citep{politisromano1994stationary}, scale-free accuracy \citep{hyndmankoehler2006mase}, and proper scoring rules for the predictive distributions \citep{gneitingraftery2007scoring,bracher2021wis}. Fourth, we run the same forecasters on synthetic markets whose latent dynamics are known, from unforecastable random walks to persistent momentum, Markov-switching regimes and regional spillovers, so that the empirical ranking can be read against the mechanisms that generate it.89We are explicit about novelty, following the verified literature review of Section~\ref{sec:lit}: each architecture is a combination of known components, and we classify each as an application, an adaptation or a new combination. The answer to the title question is largely negative: gradient boosting on hedonic and real-time market features dominates the property-level task; index changes are forecastable at one to six months through the cell's own momentum and seasonality, which a pooled tree model or an autoregression captures, while the spatial graph, regime gating and hierarchical intercepts add nothing that survives ablation; the neural sequence models are no better than the no-change forecast in a 45-origin sample and vary across seeds as much as across architectures; and at twelve months every pooled model inherits the bias of the 2021--2023 boom. Two components earn their place: the quality-adjusted state-space forecaster, stable and calibrated, and an equal-weight combination of components, which beats a learned stack.1011The paper proceeds as follows. Section~\ref{sec:lit} reviews the literature and states the gap. Section~\ref{sec:data} describes the data and the institutional context. Section~\ref{sec:targets} defines the forecasting targets and information sets. Section~\ref{sec:arch} presents the architectures and Section~\ref{sec:bench} the benchmarks. Section~\ref{sec:sim} reports the simulation study. Section~\ref{sec:framework} sets out the empirical forecasting framework. Sections~\ref{sec:results} to~\ref{sec:ablation} report the main results, spatial and regional heterogeneity, probabilistic forecasting and calibration, interpretation, and ablations and robustness. Section~\ref{sec:conclusion} concludes with limitations.12