UQO Working Papers · WP20 · v1.0
From Listing Images to Property Valuation: A Multimodal AI Framework for Automated Building Characterization and the Cost Approach
Evidence from Québec single-family listings matched to the assessment roll
- Year
- 2026
- First release
- October 2026
- Latest revision
- October 2026
- Publication status
- Working paper
Listing photographs specify the building; the cost anchor sets the value
Exterior attributes are read with 84.5–95.8% agreement against broker fields, but only 3.1% of listings show a basement within ten photographs; multimodal evidence moves the cost-to-asking-price ratio from 0.75 to 0.95 (MdAPE 30.1% to 27.9%).
Abstract
The cost approach to value requires a specification of the improvements — materials, finish grade, condition, renovations — that assessment rolls do not record and that appraisers obtain by inspection. This paper asks whether that specification can be extracted from the photographs and text of a for-sale listing by general-purpose vision–language models, and what a transparent cost indication built on it is worth. It formalises an extraction-to-valuation framework with five evidence classes (administrative, visual, proxy, text-claimed, latent), deterministic cross-image aggregation, text–image–record reconciliation, a fully sourced cost schedule and uncertainty propagation, and applies it to 1,200 Québec house listings matched within 15 m to their roll unit (14,245 photographs, three extractor tiers, 104 verified references).
Visible exterior attributes are extracted with high agreement against broker fields (roof covering 95.8%, garage type 94.1%, cladding 84.5%) but storeys agree with the roll only loosely (κ = 0.51), and the most cost-relevant interior evidence is simply not shown: within the first ten photographs, 3.1% of listings show a basement and 19.6% a bathroom, a missing-not-at-random structure that no extractor can overcome. Relative to a roll-only baseline, image and text attributes raise the median cost indication and move its ratio to the asking price from 0.75 to 0.95 (median absolute deviation 30.1% to 27.9%), mainly through condition-adjusted depreciation; the gain is real but modest, and the level of the indication depends more on the published unit-cost anchor than on anything seen in the images.
A calibration against assessor building values on five metropolitan areas transfers to a geographic hold-out and implies a hard cost near $300 per square foot, consistent with 2026 Québec industry guides and well above a Montréal hard-cost band. The ontology, prompts, schedule and code are released, and the paper is explicit about what the data cannot establish: no expert labels exist, latent defects are unobservable, and the cost indication is a screening tool, not an appraisal.
Key findings
1,200
listings matched to their roll unit, 14,245 photographs
95.8 / 94.1 / 84.5%
roof, garage and cladding agreement with broker fields
κ = 0.51
storeys agreement between image and assessment roll
3.1% / 19.6%
listings showing a basement / bathroom in ten photographs
0.75 → 0.95
cost-to-asking-price ratio, roll-only to multimodal baseline
−29.9%
median indication under a Montréal hard-cost anchor
Scientific contributions
- An extraction-to-valuation framework for the cost approach
- Five evidence classes, deterministic cross-image aggregation, text–image–record reconciliation (levels B0–B4), a fully sourced cost schedule and Monte Carlo uncertainty propagation; no price enters the indication.
- Exterior facts are read reliably and abstention works
- Against broker fields the mid-size extractor agrees on roof covering in 95.8% of informative cases, garage type 94.1% and primary cladding 84.5%, with coverage above 85%; storeys agree with the roll at κ = 0.51 because roll, broker and image count half-storeys differently.
- The binding constraint is what sellers show
- Within the first ten photographs 95.0% of listings show an exterior and 67.8% a kitchen, but only 19.6% a bathroom and 3.1% a basement — a missing-not-at-random structure that makes basement finish a text-claimed attribute and leaves condition as the images' unique contribution.
- Images move the indication through depreciation
- From the roll-only to the multimodal baseline the median depreciation falls from 0.62 to 0.38, the median indication rises from $407,545 to $514,919, the ratio to the asking price moves from 0.75 to 0.95 and the median absolute deviation from 30.1% to 27.9%; on 146 recorded sales the ratio to the sale price moves from 0.70 to 0.84.
- The unit-cost anchor dominates the evidence
- Switching from a Québec 2026 industry band to a Montréal hard-cost band moves the median indication by −29.9%, more than any attribute seen in the images; the assumed depreciation schedule is too steep for old houses and too lenient for new ones.
- Assessor calibration transfers to a geographic hold-out
- Calibrating the schedule to assessor building values on five metropolitan areas transfers to the Québec City hold-out and implies a standard hard cost of about $272–306 per square foot, inside the 2026 Québec industry band.
Methodology
Methodology (4)
Schema-constrained multimodal extraction
Per-image extraction with versioned JSON schemas on three model tiers, text extraction on masked text, confidence-weighted votes, ordinal medians and means aggregated deterministically across images, then reconciled with the roll and broker fields in five baselines B0–B4.
Transparent cost model
Indication = indexed roll land value + (replacement cost new by grade, storeys and add-ons, with soft costs and mark-ups) × (1 − depreciation from condition- and renovation-adjusted effective age); pure functions over the sourced schedule.
Evaluation of every link
Agreement, Cohen's κ and Krippendorff's α with municipality cluster bootstrap; channel information content; image-budget truncation; MNAR room-presence analysis and omission simulation; confidence calibration and selective prediction; error taxonomy.
Alignment, sensitivity and calibration
Levels B0–B4 compared with indexed assessor building values, asking prices and recorded sales; one-at-a-time sensitivity, labelled scenarios, Monte Carlo intervals (median 10–90% width 26%), and an assessor-based calibration on five metropolitan areas with a Québec City hold-out.
Research data
- Listings matched to the assessment roll
- Read-only copies (17 October 2026) of the Vrai Prix databases: 3.75 million roll units (MAMH, Données Québec), 769,000 transactions, and 77,511 live Immo-Ka house, condo and lot listings from 67 connectors; 39,278 house listings reduced to 7,745 by rules R1–R10, then a stratified sample of 1,200 by metropolitan area, age band and price quartile. Photographs are read by URL and never stored; hosts whose robots.txt disallows retrieval (RE/MAX Québec) were excluded.
- Photographs and extractions
- 14,245 photographs processed by three vision–language extractor tiers (Claude 5.5 family) through the Message Batches API under strict JSON schemas, plus text extraction on price-masked descriptions; broker fields and roll attributes serve as reference; 146 listings with a recorded sale after listing.
- Cost schedule sources
- Two published unit-cost anchors (Altus 2026 Montréal hard-cost band; Québec 2026 industry guides at $257–400 per square foot by grade), storey and attachment multipliers, add-ons and mark-ups, a Statistics Canada construction price index (18-10-0289-01) and roll land values indexed to the valuation date; every parameter sourced or flagged.
Reproducibility
- Makefile pipeline (frame, sample, batches, collect, aggregate, experiments, figures, tables, paper) with a results manifest; about 4 minutes of experiments without API calls, about US$110 of batch extraction
- Pre-registered experiment plan (14 of 16 experiments executed, 2 declared infeasible), implementation audit, results-validation integrity gate and a proposed expert annotation protocol in docs/
- pytest suite on price masking, schema constraints, aggregation rules, metrics and the cost identity; about 1,180 LaTeX macros in numbers.tex feed every number in the paper
- Ontology, prompts, cost schedule and code released; the Vrai Prix databases and photographs are not redistributed (docs/data_rights.md)
Research artifacts
Version history
| Version | Date | Description | File |
|---|---|---|---|
| v1.0current | October 2026 | Current manuscript as published on www.spboucher.ai (revised edition after the 2026-10-17 scientific review for papers 1–17). | PDF · 66 p. |