spb/wp23_uqo

1# Novelty scan: closest existing work for the six proposed WP23 architectures23Scope: for each architecture in `CLAUDE.md` §4–5, the closest published predecessors (citation keys refer4to `paper/references.bib`; metadata verified 2026-10-21, see `references_verified.json`) and what would have5to be demonstrated for the proposal to be a contribution rather than a re-labelling. Descriptions rest on6verified titles and, where retrievable, on published abstracts (availability listed in `unverified.md` §C);7where only the title is available this is stated. No novelty is claimed here; the purpose is to delimit the8existing frontier so that the manuscript's "existing / adapted / new combination / original" classification9(§13 of the protocol) can be defended.1011---1213## 1. HSHN — hierarchical spatiotemporal hedonic network1415**Closest work.** Hierarchical (partially pooled) price models are established in the index literature:16Francke & van de Minne (2017, `francke2017hierarchical`) estimate a hierarchical repeat-sales model in which17granular (municipal/segment) indices are shrunk towards a common trend, and Francke (2010, `francke2010thin`)18uses a structural time-series model for thin markets; the general machinery is Gelman & Hill (2007,19`gelmanhill2007`) and the small-area tradition (Fay & Herriot 1979, `fayherriot1979`). On the ML side,20Mayer et al. (2019, `mayer2019updating`; abstract) compare six estimators including mixed-effects21regression, random forests, gradient boosting and neural networks on 123,000 Swiss sales and two updating22schemes (moving vs. extending windows), finding boosting most accurate and robust regression least23volatile; Lorenz et al. (2023, `lorenz2023interpretable`; abstract) apply interpretable ML to hedonic rent24models; Deppner & Cajias (2024, `deppnercajias2024spatialcv`; abstract) show that random cross-validation25of ML hedonic models is biased by spatial dependence and propose spatial cross-validation; Pace & Hayunga26(2020, `pacehayunga2020residuals`; title only) test whether trees/forests extract information left in the27residuals of hedonic and spatial models. A neural hedonic model with explicit municipality-level sharing28exists in Zhang et al. (2021, `zhang2021mugrep`; abstract): MugRep builds community-level graphs and a29hierarchical, multi-task representation for appraisal. Sijp & Francke (2026, `sijpfrancke2026laplacian`;30abstract) model spatiotemporal random effects in hedonic and repeat-sales models through graph-Laplacian31priors (Bayesian or ridge), which is a parametric form of hierarchical spatiotemporal pooling.3233**What must be shown.** (i) That embedding the horizon *h* and the market state Z_mt in a single34property-level network forecasting *future* transaction prices (not contemporaneous valuation, which is what35Mayer, Lorenz, Deppner, MugRep do) yields out-of-sample gains in a strict pseudo-real-time design; (ii) that36the learned partial pooling beats the parametric alternatives it mimics — a mixed-effects hedonic model37(Mayer et al.), a hierarchical Bayesian hedonic model (Gelman & Hill) and graph-Laplacian regularisation38(Sijp & Francke) — especially in low-volume municipalities, which is the stated rationale; (iii) that the39comparison uses spatially/temporally blocked validation (Deppner & Cajias), otherwise the gain is not40credible. Without (ii) the architecture is an adaptation of known hierarchical hedonics to a neural41parameterisation.4243## 2. DSGFN — dynamic spatial graph forecasting network4445**Closest work.** The deep spatiotemporal graph models DSGFN would reuse are DCRNN (`li2018dcrnn`), STGCN46(`yu2018stgcn`), Graph WaveNet (`wu2019graphwavenet`, which learns an adaptive adjacency matrix) and AGCRN47(`bai2020agcrn`, node-adaptive parameters and data-driven graph), surveyed in `wu2021gnnsurvey`,48`jin2024urbanstgnn` and `jin2024tsgnnsurvey`. In housing economics, cross-regional spatial–temporal49dependence of prices is documented by Pace et al. (1998, 2000; `pace1998spatiotemporal`,50`pace2000spatiotemporal`, spatiotemporal autoregressions for forecasting), Holly, Pesaran & Yamagata51(2010, 2011; `hollypesaran2010`, `holly2011uk`, spatial diffusion of house prices with common factors) and52Brady (2011, `brady2011diffusion`, impulse responses of regional price diffusion). Regional index53forecasting with cross-sectional information is done by Rapach & Strauss (2009, `rapachstrauss2009`),54Gupta, Kabundi & Miller (2011, `gupta2011large`; abstract: factor-augmented VAR/BVAR and large BVAR for 2055states), Bork & Møller (2015, `borkmoller2015`, dynamic model averaging across 50 states) and, most56directly, Yeo, Stitsyuk & Choi (2025, `yeo2025deepdfvar`; title only) who combine a dynamic-factor VAR with57deep learning for regional house price index forecasting. Property-level GNN valuation models that build58graphs from transactions or POIs exist (`viana2021attention`, `das2021gsne`, `bin2019peer`,59`yang2022gcnvaluation`, `guo2026heteroroadnet`), but they value houses cross-sectionally rather than60forecasting municipal index dynamics.6162**What must be shown.** (i) That an adaptive-graph ST model on monthly municipal/submarket nodes built from63transaction data beats the econometric spatial benchmarks that encode the same idea parametrically — a64spatial-lag/STAR forecasting model (Pace et al. 2000), a FAVAR/BVAR (Gupta et al. 2011; `banbura2010bvar`)65and a DFM-based model in the spirit of Yeo et al. (2025) — under Diebold–Mariano/Giacomini–White tests and66a model confidence set; (ii) that the learned edges carry information beyond contiguity/distance67(e.g. co-movement or composition similarity), demonstrated by ablating each edge type; (iii) that gains68survive the short 2021–2026 sample, because AGCRN/Graph WaveNet were developed for thousands of69high-frequency traffic observations per node. If (i) fails, DSGFN is an application of existing traffic70models to housing, which `jin2024urbanstgnn` already anticipates as a domain.7172## 3. RAREF — regime-adaptive mixture of experts7374**Closest work.** Regime switching in house price dynamics is standard: Hamilton (1989, `hamilton1989regime`)75and Kim & Nelson (1999, `kimnelson1999statespace`) provide Markov-switching (state-space) models; Capozza,76Hendershott & Mack (2004, `capozza2004anatomy`) show that serial correlation and mean reversion of77metropolitan prices vary with market conditions; Kouwenberg & Zwinkels (2014, `kouwenbergzwinkels2014`;78title only — known to use a heterogeneous-agent switching model) forecast the US market with time-varying79behavioural regimes; Plakandaras et al. (2015, `plakandaras2015forecasting`) forecast the US real house80price index around the crisis; Bork & Møller (2015) let model weights vary over time (dynamic model81averaging), which is a soft regime mechanism. Mixtures of experts with a gating network are Jacobs et al.82(1991, `jacobs1991moe`), Jordan & Jacobs (1994, `jordanjacobs1994hme`) and sparsely gated MoE (Shazeer et83al. 2017, `shazeer2017moe`); deep switching state-space models for forecasting exist in the time-series84literature (the DSSM line, `rangapuram2018dssm`).8586**What must be shown.** (i) That regimes are inferred only from information available at the forecast87origin (no look-ahead through full-sample smoothing), and that the gating reproduces economically88interpretable states (e.g. the 2022–2023 rate-driven slowdown in Quebec); (ii) that the gain is89concentrated around regime transitions and exceeds what a Markov-switching AR, a DMA scheme (Bork &90Møller) and a time-varying-parameter model deliver; (iii) that with roughly 60–70 monthly observations per91municipality the number of experts is small enough to be estimable, with robustness to seeds and to the92choice of regime proxies. Otherwise RAREF is a known MoE applied with housing covariates.9394## 4. BHPF — Bayesian hierarchical probabilistic forecasting9596**Closest work.** Hierarchical Bayesian pooling of local price indices and hedonic effects is already in97Francke & van de Minne (2017) and Francke et al. (2022/2025, `francke2022daily`, `francke2025censustract`),98with variance priors from Gelman (2006, `gelman2006prior`); small-area estimation (Fay & Herriot 1979;99Datta & Lahiri 2000, `dattalahiri2000`; Pfeffermann 2013, `pfeffermann2013sae`) is the statistical basis.100Computation would rely on NUTS (`hoffmangelman2014nuts`), variational inference (`blei2017vi`) or INLA101(`rue2009inla`); the evaluation toolkit is Gneiting & Raftery (2007, `gneitingraftery2007scoring`),102Gneiting, Balabdaoui & Raftery (2007, `gneiting2007calibration`), Gneiting & Katzfuss (2014,103`gneitingkatzfuss2014probabilistic`) and the weighted interval score (`bracher2021wis`). Probabilistic104deep forecasters that would be the ML comparators are DeepAR (`salinas2020deepar`), TFT (`lim2021tft`),105quantile-regression networks (`taylor2000quantilenn`) and conformalized quantile regression106(`romano2019cqr`, `vovk2005conformal`). Coherent probabilistic forecasts across a municipal→regional→107provincial hierarchy are the subject of forecast reconciliation (`hyndman2011hierarchical`,108`wickramasuriya2019mint`, `athanasopoulos2024reconciliation`). The M5 uncertainty competition109(`makridakis2022m5uncertainty`) is the reference for evaluating quantile forecasts over a hierarchy.110111**What must be shown.** (i) Calibration and sharpness (PIT histograms, coverage of 50/80/95% intervals,112CRPS/WIS) in pseudo-real-time against DeepAR/TFT/CQR and against a reconciled (MinT) set of univariate113probabilistic forecasts — the latter is the natural null for "hierarchical" claims; (ii) a decomposition114of predictive variance into parameter, transaction-noise, market-dynamics and macro components that is115identified by the model rather than asserted; (iii) evidence that partial pooling improves calibration in116low-volume municipalities (where Francke & van de Minne already show gains for point indices). A Bayesian117hierarchical model that only reproduces (i) at the municipal level is an application of existing118small-area/hierarchical methods.119120## 5. QANSS — quality-adjusted neural state-space model121122**Closest work.** The decomposition "price = quality + latent market + noise" is exactly the hedonic123state-space / structural time-series index tradition: Francke (2010), Francke & van de Minne (2017),124Nagaraja, Brown & Zhao (2011, `nagaraja2011autoregressive`, autoregressive hedonic with a latent index),125Hwang & Quigley (2004, `hwangquigley2004`, selectivity, quality adjustment and mean reversion), Clapp &126Giaccotto (1992, `clappgiaccotto1992`), all built on Harvey (1989, `harvey1989`), Durbin & Koopman (2012,127`durbinkoopman2012`) and the Kalman filter (`kalman1960`). Neural state-space models in the forecasting128literature are the deep state-space model (Rangapuram et al. 2018, `rangapuram2018dssm`: an RNN emits129the parameters of a linear Gaussian state-space model per series), normalizing Kalman filters (de Bézenac130et al. 2020, `debezenac2020nkf`) and structured inference networks / deep Kalman filters (Krishnan,131Shalit & Sontag 2017, `krishnan2017dkf`). Sijp & Francke (2026) combine a hedonic model with132graph-Laplacian spatiotemporal effects, which is the parametric counterpart of a latent municipal factor133with spatial smoothing.134135**What must be shown.** (i) Identification: that the neural hedonic encoder and the latent market state136are separately identified (e.g. by constraining the encoder to be time-invariant within a window, or by137anchoring to repeat sales), otherwise "quality" absorbs market movements and the index is not138interpretable; (ii) that the latent-state forecasts beat the WP22 index forecasting benchmark (same139transactions, same vintages) and a Francke-type structural time-series model at horizons 1–12 months;140(iii) that the nonlinear transition (if any) adds value over the linear Kalman transition used by141Rangapuram et al. The "neural + Kalman" combination per se exists (DSSM); the contribution would be the142quality/market separation on transaction-level data and its forecasting value.143144## 6. QREFA — integrated architecture145146**Closest work.** Integrated multi-component architectures for real estate exist in the KDD/SIGSPATIAL147literature: MugRep (Zhang et al. 2021) combines a hedonic feature set, a transaction graph with148asynchronous spatiotemporal dependencies and a community-level graph in a multi-task hierarchical model;149HeteroRoadNet (Guo & vanden Broucke 2026; abstract) integrates roads, target locations and POIs in a150heterogeneous GNN; Viana & Barbosa (2021) use hybrid attention over neighbours. In time-series forecasting,151the Temporal Fusion Transformer (`lim2021tft`) already integrates static covariates, gating, variable152selection and quantile outputs, and Zeng et al. (2023, `zeng2023transformers`) and PatchTST153(`nie2023patchtst`) show that simple linear or patch models can beat complex transformers, which is a154warning for any "integrated" design. Forecast combination (Bates & Granger 1969, `batesgranger1969combination`;155Timmermann 2006, `timmermann2006combinations`) is the null hypothesis for any architecture that merges156component forecasts.157158**What must be shown.** (i) Component-wise ablations in which each block (hedonic encoder, spatial graph159encoder, temporal state-space, hierarchical effects, regime head, probabilistic layer) is removed and the160loss in CRPS/RMSE is reported with block-bootstrap uncertainty (`politisromano1994stationary`,161`kunsch1989jackknife`) and an MCS (`hansen2011mcs`); (ii) that the integrated model beats a simple equal- or162MinT-weighted combination of the five component models — if it does not, the integration has no incremental163value (the protocol's own rule in §5); (iii) computational cost and stability across seeds, since the164sample is small relative to the parameter count. Positioned against MugRep and TFT, the defensible claim165is at most "a new combination" unless (ii) holds with statistically significant margins.166167---168169### Cross-cutting evaluation requirements (apply to all six)170171Pseudo-real-time design with no leakage from WP22 vintages; tests from `dieboldmariano1995`,172`harvey1997equality`, `giacominiwhite2006`, `clarkmccracken2001nested` (for nested comparisons with the173random walk), `hansen2011mcs`; scale-free accuracy via MASE (`hyndmankoehler2006mase`); interpretation via174SHAP/TreeExplainer and ALE (`lundberglee2017shap`, `lundberg2020treeexplainer`, `apleyzhu2020ale`);175transaction-selection correction following `heckman1979selection`, `gatzlaffhaurin1997`, `gatzlaffhaurin1998`,176`judseaks1994selection`, `munnekeslade2000selection`; momentum/predictability benchmarks motivated by177`caseshiller1989`, `caseshiller1990forecasting`, `capozza2004anatomy`, `glaeser2014dynamics`, `head2014search`,178`guren2018momentum`, `berachaskiba2011momentum`, `titman2014dynamics`.179