UQO Working Papers · WP14 · v2.1
Hedonic House Pricing with Language-Model and Vision Features: Evidence from Quebec Real-Estate Listings
- Year
- 2026
- First release
- October 2026
- Latest revision
- October 2026
- Publication status
- Working paper
Façade photographs add more to a hedonic model than text does
On 19,123 Quebec listings two schema-constrained AI layers lift R² from 0.717 to 0.755, 3.0 of the 3.8 points coming from the façade image; the best ensemble reaches an out-of-sample R² of 0.779 with a 14.0% median absolute error.
Abstract
Hedonic house-price models are bounded by the characteristics an analyst can measure, yet much of a home's value is recorded only in the free-text description and photographs of its listing. This paper treats general-purpose AI models as a feature-engineering technology: a large language model converts each description into structured property specifications and a vision–language model converts the façade photograph into visual attributes, both under strict JSON schemas that make the outputs auditable econometric variables.
On 19,123 Quebec house listings, the two AI layers raise the R² of an already rich log-linear hedonic specification with municipality fixed effects from 0.717 to 0.755, 3.0 of the 3.8 points coming from the façade photograph, and they raise the out-of-sample R² of every estimator from OLS to gradient boosting and neural networks (best ensemble: R² 0.779, median absolute error 14.0%). The vision layer adds more to the linear model than switching from OLS to a random forest does. The extracted attributes carry economically sensible implicit prices — 2.3% per point of curb appeal, 3–4% for brick or stone façades — that survive municipality-clustered inference, a cluster bootstrap, postal-code fixed effects, and a control for the language of the description, which itself carries a 7% conditional premium for English listings.
The paper also documents that for-sale-by-owner descriptions embed the asking price, so that the text layer's gain is credible only on the brokered platform, and that spatially grouped cross-validation lowers all accuracy levels by three to five points while preserving the AI increments.
Key findings
19,123
Quebec house listings, Realtor.ca and DuProprio, July 2026
0.717 → 0.755
in-sample R² with the text and vision layers
3.0 of 3.8 pt
R² gain attributable to the façade photograph
0.779
out-of-sample R² of the best ensemble, MdAPE 14.0%
+2.3%
implicit price per point of curb appeal (1–10 scale)
19.7%
listings whose description embedded the asking price
Scientific contributions
- AI models as auditable feature engineering
- Schema-constrained decoding turns LLM and vision outputs into ordinary regressors — 18 text-derived and 21 façade-derived variables — at a marginal cost of a fraction of a cent per listing.
- The vision layer beats a change of estimator
- Adding façade attributes raises the out-of-sample R² of OLS by 5.8 points (0.628 to 0.686), more than the 4.9 points gained by switching from OLS to a random forest; gradient boosting moves from 0.743 to 0.770.
- Economically sensible implicit prices
- Curb appeal is worth 2.3% per point (cluster-bootstrap 95% CI [0.1%, 3.6%]), brick and stone façades 3.4% and 3.7%, vinyl −4.4%; the estimates hold under municipality-clustered SEs, HC3, a 299-replication cluster bootstrap and 250 postal-area fixed effects.
- Asking-price leakage quantified, not papered over
- All 3,767 DuProprio descriptions embed the platform's structured sheet including the asking price; the text layer's gain is therefore reported by platform and is credible only on Realtor.ca (+0.6 point in-sample versus +3.3 on DuProprio). The vision layer is not exposed.
- Spatially grouped validation preserves the AI increments
- Holding out entire postal areas lowers every accuracy level by three to five points (GBM 0.690 → 0.703 → 0.734) but leaves the gains of both layers intact.
- Language of the description priced
- 44.6% of descriptions are French, 43.1% English and 12.3% mixed; an English listing carries a +7.2% conditional premium within municipality, and the AI coefficients are unchanged when it is controlled for.
Methodology
Methodology (5)
Schema-constrained extraction
Strict JSON schemas with retry and backoff for both layers; text decoded at temperature zero, vision at the API default; regressors regenerable from archived outputs without API calls.
Log-linear hedonic regression with municipality fixed effects
Structural, amenity, room and listing controls plus the two AI blocks; inference by HC3, municipality-clustered standard errors (195 clusters), a pairs cluster bootstrap and postal-area (RTA) fixed effects.
Ablation by feature block and estimator
OLS, random forest, histogram gradient boosting and a residual neural network with municipality and type embeddings compared under identical random 5-fold splits and under RTA-grouped spatial folds.
Equal-weight blend with weight checks
The best ensemble averages the robust neural network and gradient boosting with a weight fixed a priori at 0.5; a weight grid (0.771–0.780) and a nested-CV weight (0.775) bracket the reported 0.779.
Indirect validation of extraction error
Agreement of LLM attributes with keyword rules (κ 0.96 for garage down to 0.50 for waterfront) and of vision attributes with text and structured fields; unreliable fields relegated to the role of controls.
Research data
- Quebec house listings
- 27,302 Realtor.ca and 7,693 DuProprio listings crawled on 3 July 2026; single-family houses and townhouses with a valid price, outliers removed, giving 19,123 observations (15,356 Realtor.ca, 3,767 DuProprio) with living area, lot, rooms, year built, type, municipality, postal code and 15 bilingual keyword amenities.
- AI feature layers
- Each description parsed by a large language model into a fixed JSON schema of specifications (renovations, finishes, systems, extras; 18 regressors) and each façade photograph parsed by a vision–language model into 12 visual fields (cladding, storeys, garage, curb appeal; 21 regressors); exteriors detected for 95.8% of listings, raw outputs archived.
Reproducibility
- End-to-end Python pipeline of 19 numbered scripts from API extraction to LaTeX tables; econometrics and ML run in about 25 minutes on an Apple-silicon laptop
- Revision analyses (language, leakage, RTA fixed effects, spatial CV, cluster bootstrap, agreement) in scripts/18 with outputs in results/revision_2026-10-17.json and tables from scripts/19
- Six figures and eleven LaTeX tables regenerated from results; full LaTeX source in paper/ with sections and bibliography (53 references, DOIs verified)
- Pre-revision version 2.0 frozen as docs/main_pre-revision-20261017.pdf with a full audit trail in docs/CHANGELOG-REVISION.md and docs/REVIEW-REPORT.md
Research artifacts
Version history
| Version | Date | Description | File |
|---|---|---|---|
| v2.1current | October 2026 | Current manuscript as published on www.spboucher.ai (revised edition after the 2026-10-17 scientific review for papers 1–17). | PDF · 46 p. |