UQO Working Papers · WP09 · v1.3
A Grand Hedonic Model of the Canadian Housing Market
Decomposing the Value of Structure and Location across 142,121 MLS Listings with High-Dimensional Neighbourhood Fixed Effects
- Year
- 2026
- First release
- October 2026
- Latest revision
- October 2026
- Publication status
- Working paper
In Canada, the neighbourhood is worth at least as much as the dwelling
Across 142,121 MLS listings, 1,155 neighbourhood fixed effects lift R² from 44% to 78%; a Shapley–Owen decomposition credits structure with 34 points and location with 36, and neighbourhood premia span a factor of about ten.
Abstract
This paper estimates a national hedonic price function for the Canadian residential market on a cross-section of 142,121 active MLS listings from nine provinces. A semi-logarithmic specification relates the log list price to structural attributes, dwelling type and ownership form, and 1,155 absorbed neighbourhood (Forward Sortation Area) fixed effects estimated by absorbing least squares with neighbourhood-clustered inference. Moving from a structural model to the full specification raises the explained share of log-price variation from 44% to 78%. Because structure and location are correlated, the attribution of this gain depends on the order in which blocks enter: a Shapley–Owen decomposition assigns about 34 percentage points to structure and 36 to location, so that the two are of comparable importance and the neighbourhood is the single largest block once province and neighbourhood are grouped.
Conditional on location, the living-area elasticity is 0.53 (0.49–0.62 across samples), each full bathroom adds about 12%, and bedroom counts are economically negligible. The implicit prices of area and bathrooms survive province-level wild-cluster bootstrap inference, whereas the small half-bathroom and bedroom effects do not. Neighbourhood effects remove 87% of the spatial autocorrelation in residuals (Moran's I from 0.50 to 0.06). The model values held-out listings in neighbourhoods seen in training with an out-of-sample R² of 0.77 and a median absolute percentage error of 15.6%; for neighbourhoods never seen in training, spatial interpolation of the neighbourhood effect still yields R² = 0.74 and a median error of 17%.
Floor space displays mild diminishing returns, the location premium falls by about 5.5% per doubling of distance to the nearest metropolitan centre, and implicit prices vary across the price distribution and across provinces. Neighbourhood premia span a factor of about ten between the most and least expensive areas. Because the dependent variable is the asking price rather than the transaction price, all estimates describe the pricing of listings, not of sales.
Key findings
142,121
active MLS listings across nine provinces
1,155
absorbed neighbourhood (FSA) fixed effects
44% → 78%
explained log-price variance, structural to full model
34 / 36 pp
Shapley–Owen shares of R² for structure and location
0.53
living-area elasticity conditional on neighbourhood
15.6%
out-of-sample MdAPE in neighbourhoods seen in training
Scientific contributions
- An order-free answer to how much location is worth
- The sequential ladder credits structure with 44 points of R² and location with 32; the Shapley–Owen decomposition gives 34 and 36, making location the largest block once province and neighbourhood are grouped.
- Stable structural implicit prices
- Living-area elasticity 0.529 (0.49–0.62 across eight sample cuts), +11.9% per full bathroom, bedrooms −0.006 and half-baths −0.034; area and bathroom effects survive province-level wild-cluster bootstrap (p = 0.004), the small effects do not.
- Neighbourhood effects absorb spatial dependence
- Residual Moran's I falls from 0.50 to 0.06 (k = 10), with the same pattern at k = 5 and k = 20; a transparent fixed-effects design removes 87% of the autocorrelation that motivates heavier spatial machinery.
- Valuation accuracy depends on spatial support
- Random 80/20 holdouts give R² 0.772 and MdAPE 15.6% (five seeds: 0.769 ± 0.010); holding out 20% of FSAs gives R² 0.74 and MdAPE 17.3% with inverse-distance interpolation of the neighbourhood effect, versus 0.52 and 25% with province means.
- Dwelling type and urban gradient
- Relative to an apartment of the same size and neighbourhood, a house lists for +46%, a triplex +43% and a duplex +16% (Kennedy-corrected); the location premium declines 5.5% per doubling of distance to the nearest of nine metros, from +21% within 5 km to −38% beyond 320 km.
- Neighbourhood premia span a factor of ten
- Vancouver's west side (V6S +226%, V6T +201%), Whistler and West Vancouver top the ranking; rural Newfoundland (A0B −69%) and Saskatchewan bottom it; the first Ontario FSA (M5R) ranks 14th.
Methodology
Methodology (5)
Absorbing least squares with a model ladder
Semi-log price on log area, structural attributes, dwelling type and ownership with M1–M5 nesting from structure to province to 1,155 FSA effects; FSA-clustered baseline inference, with HC1, province-clustered, two-way and restricted wild-cluster bootstrap (Rademacher, 512 draws) comparisons.
Shapley–Owen variance decomposition
R² attributed to structure, type, province and neighbourhood blocks by averaging over all orderings, contrasted with the sequential ladder.
Out-of-sample validation with and without spatial support
Repeated random 80/20 splits (five seeds) with Duan-smeared retransformation, and a spatial holdout of 20% of FSAs with province-mean or inverse-distance interpolation of the neighbourhood effect.
Nonlinearity, gradients and heterogeneity
Quadratic and spline floor-area terms, quantile hedonic regressions, a distance-to-metro gradient with metro intercepts and a quadratic, eight robustness cuts, nine provincial subsamples and leave-one-province-out transfer.
Spatial diagnostics
Moran's I on structural and full-model residuals with k = 5, 10 and 20 nearest neighbours and permutation p-values.
Research data
- Canadian MLS listings
- 142,121 active listings (83,524 houses, 57,853 condos, 744 other) with asking prices, drawn from a 172,019-listing Realtor.ca DuckDB snapshot; province derived from the postal-code letter, 1,146 FSAs with at least 25 listings plus nine pooled residual categories; median price $639,900 and living area 135.3 m².
Reproducibility
- Six numbered Python scripts from the committed analysis parquet (sample build, core and extended estimation, figures, tables, revision analyses); total wall-clock about 2.5 minutes with fixed seeds
- All v1.3 numbers come from results/reproduced/; the v1.2 reference outputs are kept untouched and reconciled in Appendix C of the paper
- Seven new tables (Shapley, standard-error comparison, splines, spatial holdout, Moran sensitivity, type effects, gradient) and figures generated programmatically
- Full LaTeX source of the 52-page paper; pre-revision PDF, review report and changelog archived in docs/
Research artifacts
Version history
| Version | Date | Description | File |
|---|---|---|---|
| v1.3current | October 2026 | Current manuscript as published on www.spboucher.ai (revised edition after the 2026-10-17 scientific review for papers 1–17). | PDF · 52 p. |