UQO Working Paper No. 23· v1.2· November 2026· 97 pages
Beyond Hedonic Models: Spatiotemporal, Hierarchical and Probabilistic Architectures for Residential Real Estate Price Forecasting
A pseudo-real-time evaluation on 632,000 Quebec transactions and 245 local market index series, 2021–2026
On Quebec data, no new forecasting architecture improves on gradient boosting for sale prices, and no learned model beats a 36-number calendar table of seasonal means for index changes by a detectable amount at any horizon.
Five candidate architectures and a stacked combination are compared with 8 property-level and 19 panel benchmarks on 632,006 registered sales and 245 real-time index series. The forecastable component of the index is its calendar; spatial, regime and hierarchical components add nothing that survives a seed-averaged ablation. The sample is short (13 to 33 evaluated origins per horizon, one macroeconomic cycle), so the result bounds the gain of learned models rather than ruling it out.
Simon-Pierre Boucher — contact@spboucher.ai
+0.77 [+0.36, +1.20]
Hierarchical hedonic network minus LightGBM, log points of MAE on sale prices at one month; the difference is +0.41 [−0.14, +0.99] at three months and −0.05 [−0.67, +0.52] at twelve, so the network matches the trees from three months on without beating them
21.52 [20.76, 22.30]
LightGBM MAE in log points for sales one to three months ahead, against 23.97 for the hedonic OLS; a bias of −3.7 log points appears at seven to twelve months because the market's appreciation is not extrapolated
0.75 vs 0.80
Relative MAE against no change at three months, pooled LightGBM against the seasonal mean (36 numbers); the difference of −0.19 log points has an interval [−0.44, +0.06] that includes zero
0.61
Relative MAE of the seasonal mean at twelve months, the best forecast with the calendar dummies (0.60); LightGBM reaches 0.76 and the pooled ridge is worse than no change, on only 13 origins from one episode
p = 0.012
Holm-adjusted p-value of LightGBM's gain over no change at one month, the only horizon where a learned gain survives the family-wise adjustment
−0.27 [−0.55, +0.02]
Change in six-month municipal MAE when the prior graph is removed from the spatial graph network (seed-averaged over 5 seeds): the graph carries no information
88 / 90 / 87 / 94 %
Coverage of the state-space forecaster's 90 % intervals at one, three, six and twelve months with the corrected predictive variance; nominal coverage is rejected at one and twelve months, and no model is calibrated at every horizon
Abstract
Do new forecasting architectures improve residential price forecasts over strong conventional methods? On the audited Quebec land-registry sample of UQO Working Paper 22 (632,006 sales, 2021–2026) and its real-time quality-adjusted index vintages, we forecast in a strict pseudo-real-time design the price of a dwelling sold one to twelve months ahead, the one- to twelve-month change of municipal, regional and provincial indices, and their predictive distributions, comparing five candidate architectures and a stacked combination with 8 property-level and 19 panel benchmarks. For sale prices, gradient boosting is the most accurate model at one month (hierarchical hedonic network − LightGBM +0.77 log points of MAE, bootstrap interval [+0.36, +1.20]) and the network matches it, without beating it, from three months on once every model receives the horizon as a feature; the version-1.1 comparison, whose horizon embedding was trained at four of twelve horizons, is withdrawn. For index changes, no architecture improves on a pooled tree model, and no learned model improves by a detectable amount on a 36-number calendar table of seasonal means at any horizon (relative MAE against no change 0.75 against 0.80 at three months; difference -0.19 log points, interval [-0.44, +0.06]): the forecastable component of the index is its calendar. Spatial, regime and hierarchical components add nothing that survives a seed-averaged ablation, and no model's intervals are calibrated at every horizon.
JEL: R31, C53, C45, C11, C21, C55
Key Results
Three tasks in a leakage-audited real-time design
For every monthly origin from December 2022 to September 2025 the information set contains only property characteristics, real-time municipal market features computed from past sales and the real-time index vintages of Working Paper 22, re-estimated at every publication date so that no full-sample index enters a predictor.
Architectures tested against the nulls they claim to improve on
Five candidate architectures and an integrated stack face 8 property-level and 19 panel benchmarks, from no-change, momentum and calendar rules through autoregressive, spatial-lag, factor and state-space models to gradient boosting and recurrent, convolutional and attention networks.
A negative answer for the price of a future sale
Once every model receives the horizon as a feature, gradient boosting is the most accurate model at one month and the hierarchical hedonic network is indistinguishable from it from three months on; the learned municipality embeddings add nothing beyond the municipal market features.
The forecastable component of the index is the calendar
A table of class-by-level-by-month-of-origin means is not beaten by a detectable amount by any learned model at any horizon, under final or pseudo-real-time training labels; the remaining short-horizon structure is the cell's own momentum, not spatial information.
Seed-averaged ablations and seed variability
The spatial graph, the regime gating and the hierarchical intercepts add nothing on the 64-municipality panel; neural sequence models vary across ten seeds by as much as the differences between them; an equal-weight combination of components is not worse than the learned stack.
A simulation in the same design
The same forecasters run on synthetic markets with known latent dynamics. Only a process with a calibrated annual cycle reproduces the real-data gains, so the simulation is consistent with the empirical ranking rather than an explanation of it.
Quality control
This paper went through a multi-stage quality control: 7 quality-control reports and 103 points checked. The version presented on this page is the revised manuscript. Scores are on a 0–100 scale and are a diagnostic of the manuscript, not an acceptance probability.
Score after revision
62.4/100
53.9 before quality control · +8.5
Quality control dated November 7, 2026
- Quality-control decision
- claims not supported as written → reframed and revised
- Scientific integrity
- PASS
- Reproducibility
- R2partially reproduced
- Maturity
- Bone revision cycle from submission
- Quality-control reports
- 7
- Points checked
- 103
Scores by dimension
Before quality controlAfter revision
- Originality & contribution54.0 → 56.0
- Scientific rigor48.3 → 63.0
- Identification & methodology54.0 → 64.0
- Data & reproducibility49.3 → 58.0
- Robustness & validation59.3 → 66.0
- Writing64.3 → 64.0
- Figures & tables57.1 → 72.0
- Literature positioning64.3 → 74.0
Revised manuscript97 pages13 figures58 tables1 equations142 references40,984 words
What quality control changed
Quality control found that the hierarchical network's horizon embedding had been trained at only four of the twelve evaluated horizons, so the earlier sale-price comparison was withdrawn and rerun with the horizon given to every model. Adding calendar benchmarks showed that the forecastable part of index changes is seasonality rather than anything a learned model extracts, and the claim that one model has calibrated intervals at every horizon was dropped.
- "Gradient boosting is the most accurate model at every horizon" → most accurate at one month (network − LightGBM +0.77 [+0.36, +1.20] log points); the network matches it, without beating it, from three months on; the earlier comparison is withdrawn
- "Index changes are forecastable through momentum and seasonality" (pooled tree model 0.75 at three months) → a 36-number seasonal mean reaches 0.80 and is not beaten by a detectable amount by any learned model at any horizon (difference −0.19 [−0.44, +0.06] log points)
- "At twelve months nothing beats an autoregression" (0.63) → the calendar benchmarks are the best twelve-month forecasts (seasonal mean 0.61, calendar dummies 0.60), ahead of the pooled AR(2) projection (0.63); pooled linear models fail through the linear time trend they extrapolate
- "The state-space forecaster is the only model with calibrated intervals at every horizon" (97 % at three months) → that figure came from an over-covering variance; with the corrected variance coverage is 88, 90, 87 and 94 %, and no model is calibrated at every horizon
- Ablations on a single seed → seed-averaged: removing the graph changes the six-month municipal error by −0.27 [−0.55, +0.02], an interval that now includes zero
- 14 benchmarks → 8 property-level and 19 panel benchmarks, including seasonal mean, calendar dummies, spatial-lag ridge, pooled STAR and a dynamic-factor AR; 82 cells → 245 index series
Limitations stated in the paper
- The sample is short: 69 months, 13 to 33 evaluated origins per horizon and one macroeconomic cycle; "not beaten by a detectable amount" bounds the residual gain of the learned models rather than ruling it out, and the twelve-month results describe one episode.
- The index-change target is the final smoothed index of Working Paper 22, itself an estimate whose revisions enter the target but not the real-time predictors.
- The sale-price task does not forecast which dwellings will sell; selection into transaction is documented and not corrected, and the characteristics omit bedrooms, bathrooms, condition and listing information.
- The neural architectures use fixed hyper-parameters, few epochs and, for sale prices, 3 seeds; a tuned architecture could rank differently among the neural models.
Data
Quebec land-registry sample of Working Paper 22
The snapshot of 21 October 2026 of Quebec land-registry deeds, geocoded and matched to the provincial assessment roll (769,995 registered sales from 4 January 2021 to 7 October 2026), reduced by the exclusion ledger to 632,006 house, condominium and plex sales with roll characteristics. Bedrooms, bathrooms, condition and listing attributes are not observed; no image or text features are used.
Real-time quality-adjusted index vintages
245 index series: 82 cells (the province, 17 regions and 64 municipalities) times three property classes, with 33, 31, 25 and 13 evaluated origins at one, three, six and twelve months. The target is the change of the final smoothed index.
Property-level evaluation sample
12 quarterly origins and 1,255,898 evaluated sale × origin pairs (388,780 distinct sales), by exact horizon and horizon band.
Methodology
Five architectures and an integrated stack
A hierarchical spatiotemporal hedonic network, a dynamic spatial graph forecasting network, a regime-adaptive mixture of experts, a Bayesian hierarchical probabilistic forecaster and a quality-adjusted neural state-space model, plus a stacked combination; each is classified as an application, an adaptation or a new combination of known components.
Benchmarks with distributional outputs
Hedonic OLS, random forest, XGBoost, LightGBM, support vector regression and a quantile MLP for sale prices; no-change, mean-growth, momentum, seasonal and calendar rules, pooled AR(2), spatial-lag ridge, pooled STAR, dynamic-factor AR, ridge, LightGBM, MLP, LSTM, GRU, TCN and TFT-lite for index changes. No hyper-parameter was tuned on test origins.
Inference for short samples
Diebold–Mariano tests with horizon-adjusted long-run variances, Clark–West tests for nested comparisons, Holm-adjusted families, Giacomini–White conditional tests, stationary block bootstraps with minimum detectable effects, seed-averaged comparisons for the neural models and proper scoring rules for the predictive distributions.
Simulation with known latent dynamics
Six data-generating processes on the synthetic markets of Working Paper 22: random walks, persistent momentum, Markov-switching regimes, regional spillovers, sparse cells and a calibrated annual cycle; 20 replications of 72 months, with the real-time index re-estimated on the synthetic sales at each origin.
Ablations and robustness
Component ablations as seed-averaged paired differences with origin × seed bootstrap intervals, feature-group ablations, a trend-term ablation that locates the twelve-month failure of pooled linear models, ten-seed re-estimation of the neural models and equal-weight versus learned combinations.
Reproducibility
- All code is in the repository, with a pinned environment, a root Makefile that chains panel construction, estimation, evaluation, tables, figures and the paper, and unit, chronology and tolerance-regression tests.
- The index-change analysis panel (39,690 rows) and the corresponding prediction files are committed with their checksums, and every index-change exhibit recomputes exactly from them.
- The sale-price prediction files (about 613 MB) and the simulation panels are not committed; their checksums are listed and the files are provided by the author on request, while the per-origin loss series by model and exact horizon is committed.
- The transaction data are proprietary to the provider and are not distributed with the paper; the derived inputs from Working Paper 22 are available from the author to researchers who hold a licence to the underlying feed.
- Every number quoted in the text is a generated macro, and the manuscript compiles from a clean clone.
Cite this paper
Boucher, S.-P. (2026). Beyond Hedonic Models: Spatiotemporal, Hierarchical and Probabilistic Architectures for Residential Real Estate Price Forecasting. UQO Working Paper No. 23, Université du Québec en Outaouais.
@techreport{boucher2026wp23,
author = {Simon-Pierre Boucher},
title = {Beyond Hedonic Models: Spatiotemporal, Hierarchical and Probabilistic Architectures for Residential Real Estate Price Forecasting},
institution = {Université du Québec en Outaouais, Département des sciences administratives},
number = {23},
type = {UQO Working Paper},
year = {2026},
month = {nov},
url = {https://www.spboucher.ai/research/wp23},
}More in Real Estate & Urban Economics