UQO Working Paper No. 9· v1.4· November 2026· 74 pages
A Grand Hedonic Model of the Canadian Housing Market
Structure, Location and the Dispersion of Asking Prices across 142,121 MLS Listings with High-Dimensional Neighbourhood Fixed Effects
Structure and location explain comparable shares of the dispersion of Canadian asking prices: 37.3 and 34.0 percentage points of an R² of 0.780, and neither dominates under every reasonable partition.
A single pooled hedonic equation is estimated on 142,121 active realtor.ca listings from nine provinces with 1,146 absorbed neighbourhood (FSA) effects plus 9 province-level pools. Which of structure or location ranks first depends on the geographic resolution and on where dwelling type is placed. Valuation accuracy depends on spatial support, and every estimate describes asking prices, not transaction prices.
Simon-Pierre Boucher — contact@spboucher.ai
37.3 / 34.0 pp
Grouped Shapley shares of structure and location in the grand model's R² of 0.780 (dwelling type and ownership: 6.8). The location − structure difference is −3.3 points [−5.8, −1.1], but its sign reverses when dwelling type is grouped with location (35.2 against 42.8) and when the geography is refined.
0.539
Living-area elasticity within neighbourhoods (95% CI 0.520–0.557, neighbourhood-clustered); it ranges from 0.355 to 0.633 across a specification curve of 36 alternative samples and codings, always with the same sign and p < 0.05.
+11.9%
Asking-price difference per full bathroom, conditional on neighbourhood; area and bathroom effects survive a province-level wild-cluster bootstrap, the small half-bathroom and bedroom effects do not.
+29.3% [24.9, 34.0]
Asking-price premium of a house over an apartment of the same size in the same neighbourhood, after dropping a block of listing-category dummies nested in building type that inflated the earlier estimate.
15.2% → 19.9%
Median absolute percentage error: held-out listings in trained neighbourhoods (out-of-sample R² 0.778) versus withheld contiguous blocks of neighbourhoods; in blocked hold-outs it rises from 17.4% within 5 km of a trained FSA to 31.7% beyond 50 km.
−9.5% per doubling
Change in the neighbourhood effect per doubling of distance to the nearest of nine metropolitan centres; premia net of structure span a factor of 10.6 between Vancouver's west side (V6S) and rural Newfoundland (A0B).
0.463
Mean held-out within-province R² in leave-one-province-out cross-validation with province effects in training; British Columbia, negative in the earlier version, is 0.314 once location is controlled in training.
Abstract
How much of the dispersion of Canadian asking prices is associated with the dwelling and how much with its location? We estimate a semi-logarithmic hedonic price function on 142,121 active realtor.ca listings, absorbing 1,146 Forward Sortation Area fixed effects plus 9 province-level pools. Structure and location explain comparable shares of the variance of log asking prices: a grouped Shapley decomposition assigns them 37.3 and 34.0 percentage points of an R² of 0.780, and their ranking depends on the geographic resolution and on the placement of dwelling type. Within neighbourhoods the living-area elasticity is 0.539, a dwelling is listed about 11.9% higher per full bathroom, and neighbourhood effects change by −9.5% per doubling of distance to the nearest metropolitan centre. Held-out listings in trained neighbourhoods are valued with a median absolute error of 15.2%, rising to 19.9% when contiguous blocks of neighbourhoods are withheld. The estimates describe asking prices, not transactions.
JEL: R31, R21, C21, C55
Key Results
An explicit, order-free attribution rule and its sensitivity
The explained variance is attributed by a grouped Shapley value reported under three placements of the dwelling-type block and four geographies; the location share rises mechanically with resolution, from 14.7 points with province effects to 36.7 with four-character postal effects, so the ranking of structure and location is not identified independently of the partition.
A public national listing-level hedonic equation for Canada
One pooled equation with public specification, coefficients and neighbourhood premia, set against the MLS Home Price Index, which uses sold prices but publishes neither its coefficients nor an attribution of the explained variance.
Stable structural implicit list prices
Area elasticity 0.539, about 11.9% per full bathroom, a 29.3% house premium and economically negligible bedroom counts, stable across 36 alternative fits and robust for area and bathrooms under province-level wild-cluster bootstrap inference with Rademacher and Webb weights.
Validation of the fixed-effects design in and out of spatial support
Random splits, random neighbourhood hold-outs (median 4.3 km from a trained FSA, hence interpolation inside the support; MdAPE 16.8%), spatially blocked hold-outs (19.9%) and leave-one-province-out transfer separate within-support from out-of-support accuracy.
What neighbourhood effects do to spatial dependence
The absorbed effects remove the cross-neighbourhood component of residual spatial autocorrelation; full-sample Moran's I of the grand-model residual is 0.284 with ten nearest neighbours, almost all within-FSA correlation among listings a median of 0.32 km apart, with no residual correlation beyond one kilometre.
Peer review
This paper went through a multi-referee review: 6 referee reports and 69 review items. 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
69.3/100
60.3 before review · +9.0
Review dated November 3, 2026
- Review decision
- major revision
- Scientific integrity
- PASS
- Reproducibility
- R1reproduced from documented analysis data
- Maturity
- Bone revision cycle from submission
- Referee reports
- 6
- Review items
- 69
Scores by dimension
Before reviewAfter revision
- Originality & contribution58.3 → 60.0
- Scientific rigor61.5 → 70.0
- Identification & methodology60.3 → 68.0
- Data & reproducibility54.2 → 72.0
- Robustness & validation61.8 → 74.0
- Writing71.0 → 76.0
- Figures & tables67.5 → 80.0
- Literature positioning60.0 → 74.0
Revised manuscript74 pages17 figures24 tables6 equations110 references28,947 words
What the review changed
Peer review replaced the claim that location is worth at least as much as the dwelling with a bounded statement: the two explain comparable shares of asking-price dispersion and their ranking depends on the partition. It also corrected an inflated house premium, re-estimated the urban gradient on the neighbourhood effects, and separated within-support from out-of-support valuation accuracy.
- House premium over an apartment: +46% → +29.3% [24.9, 34.0], after dropping listing-category dummies nested in building type.
- Variance decomposition: a single Shapley split (34 structure / 36 location) → reported under several partitions (37.3 / 6.8 / 34.0 with type as its own group; location − structure −3.3 [−5.8, −1.1]), with the ranking stated as partition-dependent.
- Urban gradient: −5.5% per doubling of distance (on residuals of a model without location effects) → −9.5% per doubling, estimated on the neighbourhood fixed effects.
- Unseen-neighbourhood accuracy: the random neighbourhood hold-out is relabelled as interpolation inside the support (median 4.3 km from a trained FSA, MdAPE 16.8%) and spatially blocked hold-outs are added (MdAPE 19.9%, rising with distance).
- Leave-one-province-out transfer for British Columbia: negative R² → 0.314 with province effects in training (mean across provinces 0.463).
- Spatial autocorrelation: the claim that neighbourhood effects remove 87% of it is withdrawn; the revised paper reports a full-sample Moran's I of 0.284, almost all within-FSA, with no residual correlation beyond one kilometre.
Limitations stated in the paper
- Prices are asking prices from a stock of active listings; the wedge between implicit list prices and implicit transaction prices is unknown for Canadian housing, and slow-selling dwellings are over-represented.
- The data lack year of construction, renovation status and interior quality, so within-neighbourhood estimates remain exposed to sorting, the bathroom coefficient most of all; several fields are noisy and a third of living areas are banded.
- The 1% trims are selection on the dependent variable, and the analysis is a single cross-section of the spring of 2026 that cannot speak to price dynamics.
- FSAs are postal conveniences rather than planning units, the raw file from a single listing platform cannot be redistributed, and its coverage of rural areas, private sales and new construction is unknown.
Data
realtor.ca active listings, snapshot of 2026-05-29
A file of 172,019 for-sale listings with distinct MLS numbers; 90.2% were inserted in 2026, so the cross-section is the stock of active listings in the spring of 2026. Each record carries the asking price, semi-structured building and lot attributes, dwelling type, ownership and a geocoded location.
Estimation sample
147,236 listings with price, living area, bedrooms and bathrooms, reduced to 142,121 dwellings after 1% trims on price and living area (83,524 in the house subsample), across nine provinces; Prince Edward Island listings carry no living-area field and the territories are absent. Median asking price $639,900, median living area 135.3 m²; 34.3% of living areas are banded.
Neighbourhood geography
Forward Sortation Areas (first three characters of the postal code): 1,667 distinct FSAs, those with fewer than 25 listings pooled by province, giving 1,155 absorbed location effects (1,146 FSAs and 9 pools). Province is derived from the postal letter because the collector's regional seed disagrees with it for 11.7% of the sample.
Methodology
Semi-log hedonic with absorbed neighbourhood effects
Log asking price on log living area, bedrooms, bathrooms, storeys, parking, lot, dwelling type and ownership, in a ladder M1–M5 from structure only to FSA effects on the full residential sample, estimated by absorbing least squares; Kennedy-corrected percentage effects for indicators.
Grouped Shapley decomposition under several partitions
Shapley value of the quotient game over structure, type/ownership and location (province and FSA grouped), with three placements of the type block, four geographies and a neighbourhood-cluster bootstrap (200 draws) on the location − structure difference.
Inference ladder
HC1, FSA-clustered (baseline, 1,155 clusters), two-character-prefix-clustered (127 clusters), province-clustered and bias-reduced province-level variance estimators, effective numbers of clusters, and a restricted province-level wild-cluster bootstrap-t with Rademacher and Webb six-point weights.
Four validation designs
Random 80/20 splits over five seeds; random hold-out of 20% of FSAs with imputed neighbourhood effects; spatially blocked hold-outs of two-character postal prefixes or k-means clusters with a 20 km buffer, reported by distance band; and leave-one-province-out cross-validation under three training designs. Levels are retransformed with Duan smearing.
Functional form, heterogeneity, gradient and spatial diagnostics
Quadratic and spline terms in floor space, quantile regressions with FSA-cluster bootstrap standard errors, a distance gradient estimated on the FSA fixed effects for nine metropolitan centres, and Moran's I split into within- and cross-FSA parts with a distance-band correlogram.
Reproducibility
- The repository holds the processed estimation sample, every numbered script of the pipeline and the full LaTeX source of the paper.
- A single entry point (run_all.sh) rebuilds all tables and figures from the processed sample and writes a run log with package versions, seeds and the count ladder; a test suite compares regenerated outputs with the archived ones.
- Every number quoted in the text is a macro generated from the result files, and an appendix reconciles the estimates with the May and October 2026 versions.
- The raw listing file is not redistributed because the portal's terms of use restrict redistribution; its size, column count and SHA-256 checksum are recorded in the paper so that a copy can be authenticated.
Cite this paper
Boucher, S.-P. (2026). A Grand Hedonic Model of the Canadian Housing Market. UQO Working Paper No. 9, Université du Québec en Outaouais.
@techreport{boucher2026wp9,
author = {Simon-Pierre Boucher},
title = {A Grand Hedonic Model of the Canadian Housing Market},
institution = {Université du Québec en Outaouais, Département des sciences administratives},
number = {9},
type = {UQO Working Paper},
year = {2026},
month = {nov},
url = {https://www.spboucher.ai/research/wp9},
}More in Real Estate & Urban Economics