UQO Working Papers · WP23 · v1.1
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 82 local market indices, 2021–2026
- Year
- 2026
- First release
- October 2026
- Latest revision
- October 2026
- Publication status
- Working paper
On Quebec data, new forecasting architectures do not beat gradient boosting
LightGBM reaches a 21.21 log-point MAE on future sale prices against 22.36 for the hierarchical network; for index changes a pooled tree model attains MASE 0.75 at three months and nothing beats an autoregression at twelve.
Abstract
This paper asks whether new forecasting architectures improve residential price forecasts over strong conventional methods, and under which conditions. Using the audited Quebec land-registry sample of UQO Working Paper 22 (632,006 sales, 2021–2026) and its real-time quality-adjusted index vintages, it defines three tasks in a strict pseudo-real-time design: the price of a property sold one to twelve months after a forecast origin (Task A), the one- to twelve-month change of municipal, regional and provincial quality-adjusted indices (Task B), and calibrated predictive distributions for both (Task C).
Five candidate architectures are implemented (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) together with an integrated stacked architecture, against 14 benchmarks ranging from no-change and momentum rules to gradient boosting, recurrent, convolutional and attention networks. Evaluation uses Diebold–Mariano and Giacomini–White tests, model confidence sets and block bootstraps, with Monte Carlo experiments on synthetic markets whose latent dynamics range from unforecastable random walks to regime switching and regional spillovers.
For the price of a future sale, gradient boosting on hedonic and real-time local market features is the most accurate model at every horizon and the hierarchical network does not improve on it. For index changes, one- to six-month movements are forecastable through the cell's own momentum and seasonality (MASE 0.75 at three months for a pooled tree model); spatial, regime and hierarchical components add nothing that survives ablation, and at twelve months nothing beats an autoregression. The state-space forecaster is the only model with calibrated intervals at every horizon. Every component of the architectures exists in the literature; the contribution is the measured answer, on transaction-level data, to the question of when complexity pays.
Key findings
632,006
Quebec registry sales in the audited sample, 2021–2026
6 vs 14
candidate architectures against benchmark forecasters
21.21
LightGBM log-point MAE on sale prices, 1–3 months
0.75
best MASE for three-month index changes, pooled tree model
0.63
AR(2) MASE at twelve months, unbeaten by any architecture
97%
coverage of state-space 90% intervals at three months
Scientific contributions
- 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 local market features and the WP22 index vintages re-estimated at each publication date; no full-sample index enters a predictor.
- Gradient boosting dominates the property-level task
- At one to three months LightGBM reaches a 21.21 log-point MAE against 22.36 for the hierarchical network, 23.24 for hedonic OLS and 34.96 for the municipal median; at seven to twelve months every model is biased by −4 to −9 log points because the market is not extrapolated.
- Index changes are forecastable only through momentum
- At three months a pooled LightGBM reaches MASE 0.75 versus no change, with the gain concentrated in rising windows (0.63 against 1.00 in flat ones); after Holm adjustment no improvement survives a 10% family-wise error rate.
- Spatial, regime and hierarchical components add nothing
- Removing the graph from the spatial network lowers its municipal six-month error by 0.59 log points, an equal-weight combination beats the learned stack by 0.38, and removing momentum from ridge raises error by 0.53; neural sequence models vary across seeds as much as across architectures.
- At twelve months, nothing beats an autoregression
- AR(2) attains MASE 0.63 while every pooled model is worse than no change, inheriting a +6 to +11 log-point bias from the 2021–2023 boom.
- Only the state-space forecaster is calibrated
- The quality-adjusted neural state-space model is the only architecture with honest intervals at every horizon (97% coverage of 90% intervals at three months, against 73–76% for LightGBM, the graph network and the attention benchmark).
Methodology
Methodology (5)
Five architectures and an integrated stack
Hierarchical spatiotemporal hedonic network, dynamic spatial graph forecasting network, regime-adaptive mixture of experts, Bayesian hierarchical probabilistic forecaster, quality-adjusted neural state-space model, and a stacked combination, each classified as an application, adaptation or new combination of known components.
Fourteen benchmarks with distributional outputs
No-change and momentum rules, ARIMA-type, factor and state-space models, ridge, LightGBM, XGBoost, random forest, SVR, MLP, and LSTM, GRU, TCN and TFT-lite networks, each with a quantile or distributional output.
Forward-chaining inference for short samples
Diebold–Mariano tests with horizon-adjusted long-run variances, Giacomini–White conditional tests, model confidence sets, stationary block bootstraps, scale-free MASE and proper scoring rules (CRPS, interval scores, Brier skill).
Simulation with known latent dynamics
The same forecasters run on synthetic markets built with the WP22 simulator: random walks, persistent momentum, Markov-switching regimes, regional spillovers and sparse cells, so the empirical ranking can be read against the generating mechanism.
Ablations and robustness
Component ablations, seed variability, ex-post regime splits, directional accuracy at turning points, municipal-cell distributions and equal-weight versus learned combinations.
Research data
- WP22 Quebec registry sample
- 632,006 registered sales from January 2021 to 7 October 2026 with assessment-roll characteristics (no listing, bedroom, bathroom, condition or image data), read-only from the SHARP-HPI repository; forecast origins span December 2022 to September 2025 with 31–45 origins per horizon.
- Real-time quality-adjusted index vintages
- SHARP-HPI vintages published from July 2022 to October 2026 (filtered and smoothed modes) for 82 cells (province, 17 regions, 64 municipalities) and 192 municipal cells by property class, with the retrospective final index as the Task B target.
Reproducibility
- Numbered pipeline scripts/10–60 (Task A, Task B, evaluation, interpretation, ablations, robustness, simulation, tables, figures) with src/wp23 modules and a pytest suite; heavy runs executed on a 32-core Mac Studio with the same code
- Every number in the manuscript is a generated macro (paper/tables/numbers.tex); results stored per task, evaluation, ablation, robustness and simulation
- Leakage audit (docs/leakage_audit.md), decision log, limitations and a verified literature review with novelty scan (152 references)
- Full LaTeX source of the 56-page paper; inputs are read-only WP22 artefacts (raw microdata not redistributed)
Research artifacts
Version history
| Version | Date | Description | File |
|---|---|---|---|
| v1.1current | October 2026 | Current manuscript as published on www.spboucher.ai (revised edition after the 2026-10-17 scientific review for papers 1–17). | PDF · 56 p. |