# Novelty scan: closest existing work for the six proposed WP23 architectures

Scope: for each architecture in CLAUDE.md §4–5, the closest published predecessors (citation keys refer to paper/references.bib; metadata verified 2026-10-21, see references_verified.json) and what would have to be demonstrated for the proposal to be a contribution rather than a re-labelling. Descriptions rest on verified titles and, where retrievable, on published abstracts (availability listed in unverified.md §C); where only the title is available this is stated. No novelty is claimed here; the purpose is to delimit the existing frontier so that the manuscript's "existing / adapted / new combination / original" classification (§13 of the protocol) can be defended.

# 1. HSHN — hierarchical spatiotemporal hedonic network

Closest work. Hierarchical (partially pooled) price models are established in the index literature: Francke & van de Minne (2017, francke2017hierarchical) estimate a hierarchical repeat-sales model in which granular (municipal/segment) indices are shrunk towards a common trend, and Francke (2010, francke2010thin) uses a structural time-series model for thin markets; the general machinery is Gelman & Hill (2007, gelmanhill2007) and the small-area tradition (Fay & Herriot 1979, fayherriot1979). On the ML side, Mayer et al. (2019, mayer2019updating; abstract) compare six estimators including mixed-effects regression, random forests, gradient boosting and neural networks on 123,000 Swiss sales and two updating schemes (moving vs. extending windows), finding boosting most accurate and robust regression least volatile; Lorenz et al. (2023, lorenz2023interpretable; abstract) apply interpretable ML to hedonic rent models; Deppner & Cajias (2024, deppnercajias2024spatialcv; abstract) show that random cross-validation of ML hedonic models is biased by spatial dependence and propose spatial cross-validation; Pace & Hayunga (2020, pacehayunga2020residuals; title only) test whether trees/forests extract information left in the residuals of hedonic and spatial models. A neural hedonic model with explicit municipality-level sharing exists in Zhang et al. (2021, zhang2021mugrep; abstract): MugRep builds community-level graphs and a hierarchical, multi-task representation for appraisal. Sijp & Francke (2026, sijpfrancke2026laplacian; abstract) model spatiotemporal random effects in hedonic and repeat-sales models through graph-Laplacian priors (Bayesian or ridge), which is a parametric form of hierarchical spatiotemporal pooling.

What must be shown. (i) That embedding the horizon h and the market state Z_mt in a single property-level network forecasting future transaction prices (not contemporaneous valuation, which is what Mayer, Lorenz, Deppner, MugRep do) yields out-of-sample gains in a strict pseudo-real-time design; (ii) that the learned partial pooling beats the parametric alternatives it mimics — a mixed-effects hedonic model (Mayer et al.), a hierarchical Bayesian hedonic model (Gelman & Hill) and graph-Laplacian regularisation (Sijp & Francke) — especially in low-volume municipalities, which is the stated rationale; (iii) that the comparison uses spatially/temporally blocked validation (Deppner & Cajias), otherwise the gain is not credible. Without (ii) the architecture is an adaptation of known hierarchical hedonics to a neural parameterisation.

# 2. DSGFN — dynamic spatial graph forecasting network

Closest work. The deep spatiotemporal graph models DSGFN would reuse are DCRNN (li2018dcrnn), STGCN (yu2018stgcn), Graph WaveNet (wu2019graphwavenet, which learns an adaptive adjacency matrix) and AGCRN (bai2020agcrn, node-adaptive parameters and data-driven graph), surveyed in wu2021gnnsurvey, jin2024urbanstgnn and jin2024tsgnnsurvey. In housing economics, cross-regional spatial–temporal dependence of prices is documented by Pace et al. (1998, 2000; pace1998spatiotemporal, pace2000spatiotemporal, spatiotemporal autoregressions for forecasting), Holly, Pesaran & Yamagata (2010, 2011; hollypesaran2010, holly2011uk, spatial diffusion of house prices with common factors) and Brady (2011, brady2011diffusion, impulse responses of regional price diffusion). Regional index forecasting with cross-sectional information is done by Rapach & Strauss (2009, rapachstrauss2009), Gupta, Kabundi & Miller (2011, gupta2011large; abstract: factor-augmented VAR/BVAR and large BVAR for 20 states), Bork & Møller (2015, borkmoller2015, dynamic model averaging across 50 states) and, most directly, Yeo, Stitsyuk & Choi (2025, yeo2025deepdfvar; title only) who combine a dynamic-factor VAR with deep learning for regional house price index forecasting. Property-level GNN valuation models that build graphs from transactions or POIs exist (viana2021attention, das2021gsne, bin2019peer, yang2022gcnvaluation, guo2026heteroroadnet), but they value houses cross-sectionally rather than forecasting municipal index dynamics.

What must be shown. (i) That an adaptive-graph ST model on monthly municipal/submarket nodes built from transaction data beats the econometric spatial benchmarks that encode the same idea parametrically — a spatial-lag/STAR forecasting model (Pace et al. 2000), a FAVAR/BVAR (Gupta et al. 2011; banbura2010bvar) and a DFM-based model in the spirit of Yeo et al. (2025) — under Diebold–Mariano/Giacomini–White tests and a model confidence set; (ii) that the learned edges carry information beyond contiguity/distance (e.g. co-movement or composition similarity), demonstrated by ablating each edge type; (iii) that gains survive the short 2021–2026 sample, because AGCRN/Graph WaveNet were developed for thousands of high-frequency traffic observations per node. If (i) fails, DSGFN is an application of existing traffic models to housing, which jin2024urbanstgnn already anticipates as a domain.

# 3. RAREF — regime-adaptive mixture of experts

Closest work. Regime switching in house price dynamics is standard: Hamilton (1989, hamilton1989regime) and Kim & Nelson (1999, kimnelson1999statespace) provide Markov-switching (state-space) models; Capozza, Hendershott & Mack (2004, capozza2004anatomy) show that serial correlation and mean reversion of metropolitan prices vary with market conditions; Kouwenberg & Zwinkels (2014, kouwenbergzwinkels2014; title only — known to use a heterogeneous-agent switching model) forecast the US market with time-varying behavioural regimes; Plakandaras et al. (2015, plakandaras2015forecasting) forecast the US real house price index around the crisis; Bork & Møller (2015) let model weights vary over time (dynamic model averaging), which is a soft regime mechanism. Mixtures of experts with a gating network are Jacobs et al. (1991, jacobs1991moe), Jordan & Jacobs (1994, jordanjacobs1994hme) and sparsely gated MoE (Shazeer et al. 2017, shazeer2017moe); deep switching state-space models for forecasting exist in the time-series literature (the DSSM line, rangapuram2018dssm).

What must be shown. (i) That regimes are inferred only from information available at the forecast origin (no look-ahead through full-sample smoothing), and that the gating reproduces economically interpretable states (e.g. the 2022–2023 rate-driven slowdown in Quebec); (ii) that the gain is concentrated around regime transitions and exceeds what a Markov-switching AR, a DMA scheme (Bork & Møller) and a time-varying-parameter model deliver; (iii) that with roughly 60–70 monthly observations per municipality the number of experts is small enough to be estimable, with robustness to seeds and to the choice of regime proxies. Otherwise RAREF is a known MoE applied with housing covariates.

# 4. BHPF — Bayesian hierarchical probabilistic forecasting

Closest work. Hierarchical Bayesian pooling of local price indices and hedonic effects is already in Francke & van de Minne (2017) and Francke et al. (2022/2025, francke2022daily, francke2025censustract), with variance priors from Gelman (2006, gelman2006prior); small-area estimation (Fay & Herriot 1979; Datta & Lahiri 2000, dattalahiri2000; Pfeffermann 2013, pfeffermann2013sae) is the statistical basis. Computation would rely on NUTS (hoffmangelman2014nuts), variational inference (blei2017vi) or INLA (rue2009inla); the evaluation toolkit is Gneiting & Raftery (2007, gneitingraftery2007scoring), Gneiting, Balabdaoui & Raftery (2007, gneiting2007calibration), Gneiting & Katzfuss (2014, gneitingkatzfuss2014probabilistic) and the weighted interval score (bracher2021wis). Probabilistic deep forecasters that would be the ML comparators are DeepAR (salinas2020deepar), TFT (lim2021tft), quantile-regression networks (taylor2000quantilenn) and conformalized quantile regression (romano2019cqr, vovk2005conformal). Coherent probabilistic forecasts across a municipal→regional→ provincial hierarchy are the subject of forecast reconciliation (hyndman2011hierarchical, wickramasuriya2019mint, athanasopoulos2024reconciliation). The M5 uncertainty competition (makridakis2022m5uncertainty) is the reference for evaluating quantile forecasts over a hierarchy.

What must be shown. (i) Calibration and sharpness (PIT histograms, coverage of 50/80/95% intervals, CRPS/WIS) in pseudo-real-time against DeepAR/TFT/CQR and against a reconciled (MinT) set of univariate probabilistic forecasts — the latter is the natural null for "hierarchical" claims; (ii) a decomposition of predictive variance into parameter, transaction-noise, market-dynamics and macro components that is identified by the model rather than asserted; (iii) evidence that partial pooling improves calibration in low-volume municipalities (where Francke & van de Minne already show gains for point indices). A Bayesian hierarchical model that only reproduces (i) at the municipal level is an application of existing small-area/hierarchical methods.

# 5. QANSS — quality-adjusted neural state-space model

Closest work. The decomposition "price = quality + latent market + noise" is exactly the hedonic state-space / structural time-series index tradition: Francke (2010), Francke & van de Minne (2017), Nagaraja, Brown & Zhao (2011, nagaraja2011autoregressive, autoregressive hedonic with a latent index), Hwang & Quigley (2004, hwangquigley2004, selectivity, quality adjustment and mean reversion), Clapp & Giaccotto (1992, clappgiaccotto1992), all built on Harvey (1989, harvey1989), Durbin & Koopman (2012, durbinkoopman2012) and the Kalman filter (kalman1960). Neural state-space models in the forecasting literature are the deep state-space model (Rangapuram et al. 2018, rangapuram2018dssm: an RNN emits the parameters of a linear Gaussian state-space model per series), normalizing Kalman filters (de Bézenac et al. 2020, debezenac2020nkf) and structured inference networks / deep Kalman filters (Krishnan, Shalit & Sontag 2017, krishnan2017dkf). Sijp & Francke (2026) combine a hedonic model with graph-Laplacian spatiotemporal effects, which is the parametric counterpart of a latent municipal factor with spatial smoothing.

What must be shown. (i) Identification: that the neural hedonic encoder and the latent market state are separately identified (e.g. by constraining the encoder to be time-invariant within a window, or by anchoring to repeat sales), otherwise "quality" absorbs market movements and the index is not interpretable; (ii) that the latent-state forecasts beat the WP22 index forecasting benchmark (same transactions, same vintages) and a Francke-type structural time-series model at horizons 1–12 months; (iii) that the nonlinear transition (if any) adds value over the linear Kalman transition used by Rangapuram et al. The "neural + Kalman" combination per se exists (DSSM); the contribution would be the quality/market separation on transaction-level data and its forecasting value.

# 6. QREFA — integrated architecture

Closest work. Integrated multi-component architectures for real estate exist in the KDD/SIGSPATIAL literature: MugRep (Zhang et al. 2021) combines a hedonic feature set, a transaction graph with asynchronous spatiotemporal dependencies and a community-level graph in a multi-task hierarchical model; HeteroRoadNet (Guo & vanden Broucke 2026; abstract) integrates roads, target locations and POIs in a heterogeneous GNN; Viana & Barbosa (2021) use hybrid attention over neighbours. In time-series forecasting, the Temporal Fusion Transformer (lim2021tft) already integrates static covariates, gating, variable selection and quantile outputs, and Zeng et al. (2023, zeng2023transformers) and PatchTST (nie2023patchtst) show that simple linear or patch models can beat complex transformers, which is a warning for any "integrated" design. Forecast combination (Bates & Granger 1969, batesgranger1969combination; Timmermann 2006, timmermann2006combinations) is the null hypothesis for any architecture that merges component forecasts.

What must be shown. (i) Component-wise ablations in which each block (hedonic encoder, spatial graph encoder, temporal state-space, hierarchical effects, regime head, probabilistic layer) is removed and the loss in CRPS/RMSE is reported with block-bootstrap uncertainty (politisromano1994stationary, kunsch1989jackknife) and an MCS (hansen2011mcs); (ii) that the integrated model beats a simple equal- or MinT-weighted combination of the five component models — if it does not, the integration has no incremental value (the protocol's own rule in §5); (iii) computational cost and stability across seeds, since the sample is small relative to the parameter count. Positioned against MugRep and TFT, the defensible claim is at most "a new combination" unless (ii) holds with statistically significant margins.

# Cross-cutting evaluation requirements (apply to all six)

Pseudo-real-time design with no leakage from WP22 vintages; tests from dieboldmariano1995, harvey1997equality, giacominiwhite2006, clarkmccracken2001nested (for nested comparisons with the random walk), hansen2011mcs; scale-free accuracy via MASE (hyndmankoehler2006mase); interpretation via SHAP/TreeExplainer and ALE (lundberglee2017shap, lundberg2020treeexplainer, apleyzhu2020ale); transaction-selection correction following heckman1979selection, gatzlaffhaurin1997, gatzlaffhaurin1998, judseaks1994selection, munnekeslade2000selection; momentum/predictability benchmarks motivated by caseshiller1989, caseshiller1990forecasting, capozza2004anatomy, glaeser2014dynamics, head2014search, guren2018momentum, berachaskiba2011momentum, titman2014dynamics.