UQO Working Paper No. 20· v1.1· November 2026· 102 pages

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

Real Estate & Urban Economics

On SSRN

Listing photographs yield a reproducible, extractor-specific specification of the building, but the paired gain in price alignment is −2.3 points [−5.7, +1.5] and a one-constant roll-only schedule does better.

On 1,200 Québec house listings matched within 15 m to their roll unit, judgment attributes are reproducible within a model (test–retest κ 0.77–0.95) but not across models (κ 0.49–0.63), and only 3.1% of listings show a basement within ten photographs. The extracted attributes move the median ratio of the cost indication to the asking price from 0.75 to 0.95, almost entirely through a lower effective age. This is a measurement paper: there are no expert labels, and the design does not identify valuation information in listing media beyond a correction of the age–life schedule.

Simon-Pierre Boucher — contact@spboucher.ai

Paper PDF (revised)Read online SSRN Code & data repository

95.8 / 94.1 / 84.5%

agreement of the mid-size extractor with broker fields on roof covering, garage type and primary cladding, in informative cases, with coverage above 85% for each; storeys agree with the roll only loosely (κ = 0.52)

κ 0.77–0.95 vs 0.49–0.63

agreement on judgment attributes (finish grades, renovation evidence, condition) when the same model is re-run on the same photographs versus between two models: a difference of scale, not sampling noise

3.1% / 19.6%

listings showing a basement (2.1% design-weighted) and a bathroom within the first ten photographs, against 67.8% for a kitchen and 95.0% for an exterior

0.75 → 0.95

median ratio of the cost indication to the asking price, roll-only versus multimodal baseline (0.70 to 0.91 design-weighted); median depreciation falls from 0.62 to 0.38. A level effect, almost entirely through a lower effective age

−2.3 pt [−5.7, +1.5]

paired difference in the median absolute deviation from asking prices between the multimodal and the roll-only baseline: the interval includes zero, and the difference reverses sign across age bands

+5.2 pt [+3.6, +6.8]

amount by which the multimodal indication's deviation exceeds that of a roll-only schedule re-centred on the extractors' average condition; a random reassignment of extracted attributes across listings does not raise the deviation (one-sided p = 0.95)

−29.9%

change in the median indication when the published unit-cost anchor is switched, more than any channel; the joint age-moment fit against assessor values implies a hard cost of $167 per square foot [154, 182] on a two-storey basis

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. We ask whether that specification can be extracted from the photographs and text of a for-sale listing by general-purpose vision–language models, how reliably, and what a transparent cost indication built on it is worth. We formalise an extraction-to-valuation chain with five evidence classes, deterministic aggregation and reconciliation rules, a sourced cost schedule and propagated uncertainty, and apply it to 1,200 Québec house listings matched within 15 m to their roll unit (14,245 photographs; two full-sample image tiers plus a third tier on a 200-listing subsample and on text). The contribution is a measurement. Visible exterior attributes agree with broker fields (roof covering 95.8%, garage type 94.1%, cladding 84.5%); storeys agree with the roll only loosely (κ = 0.52); judgment attributes are reproducible within a model (test–retest κ 0.77–0.95) but not across models (κ 0.49–0.63), a difference of scale rather than sampling noise. The most cost-relevant interior evidence is systematically missing by room and presentation order: within ten photographs, 3.1% of listings (2.1% design-weighted) show a basement and 19.6% a bathroom. Relative to a roll-only baseline, the extracted attributes move the median ratio of the indication to the asking price from 0.75 to 0.95 (0.70 to 0.91 weighted), almost entirely through a lower effective age; the paired difference in the median absolute deviation from asking prices is -2.3 points [-5.7, +1.5], and a roll-only schedule re-centred on the extractors’ average condition reaches a lower deviation than the multimodal indication (+5.2 points [+3.6, +6.8]). The design therefore does not establish that listing media add valuation information beyond a one-constant correction of the age–life schedule; the level of the indication is set by the published unit cost, and the calibration against assessor values is reported with its identification limits. Ontology, prompts, schedule, code and raw model responses are released.

JEL: R31, R32, C45, C55, D83

Key Results

A formal extraction-to-valuation chain for the cost approach

A cost identity with explicit scope for each term, an attribute ontology with five evidence classes (administrative, visual, proxy, text-claimed, latent), deterministic cross-image aggregation, text–image–record reconciliation, a sourced cost schedule and a condition-adjusted age–life depreciation. No parameter is fitted on prices.

Visible exterior facts are read reliably and abstention works

Roof covering, garage type and cladding agree with broker fields in the mid-eighties to mid-nineties; storeys agree loosely because the roll, the broker and the image count half-storeys differently.

Judgment attributes are an extractor-specific scale

Grades and renovation evidence are reproducible within a model and not across models, and a paraphrased prompt moves the attributes about as much as a re-run of the same prompt; what the images yield is a consistent scale rather than a measurement of condition.

The binding constraint is what sellers show

Interior evidence is systematically missing by room and presentation order, so basement finish, one of the largest add-ons in a cost schedule, is essentially a text-claimed attribute, and interior condition is judged from living rooms and kitchens.

A level effect, not identified information content

The shift of the indication is reported next to one-constant baselines: the paired difference includes zero, the deviation from the assessor's value is larger under the multimodal baseline (+4.3 points [+1.1, +7.8]), and a roll-only schedule with one constant reproduces the shift with equal or lower dispersion.

The unit-cost anchor sets the level, and calibration has identification limits

A single pooled moment does not identify the unit-cost level separately from the depreciation slope; the joint fit on five age-band medians places the hard cost well below the published anchor and the economic life far above the assumed one.

Quality control

This paper went through a multi-stage quality control: 7 quality-control reports and 107 points checked. 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

70.9/100

59.8 before quality control · +11.1

Quality control dated November 7, 2026

Quality-control decision: claims not supported as written → reframed and revised
Scientific integrity: PASS
Reproducibility: R1reproduced from documented analysis data
Maturity: Bone revision cycle from submission
Quality-control reports: 7
Points checked: 107

Scores by dimension

Before quality controlAfter revision

Originality & contribution69.1 → 66.0
Scientific rigor55.4 → 70.0
Identification & methodology50.4 → 68.0
Data & reproducibility57.9 → 76.0
Robustness & validation58.6 → 76.0
Writing72.4 → 72.0
Figures & tables64.6 → 72.0
Literature positioning67.6 → 78.0

Revised manuscript102 pages19 figures54 tables7 equations112 references41,689 words

What quality control changed

Quality control found that the original headline, a modest but real valuation gain from listing images, was not identified by the design. The revised paper is a measurement paper: it reports what is shown, what is read and how reliably, and states that listing media are not shown to add valuation information beyond a one-constant correction of the depreciation schedule.

Valuation gain: ‘the gain is real but modest’ (deviation from asking prices 30.1% → 27.9%) → paired difference of −2.3 points [−5.7, +1.5], which includes zero; a roll-only schedule re-centred on the extractors' average condition reaches a lower deviation (+5.2 points [+3.6, +6.8]).
Implied construction cost: a hard cost ‘near $300 per square foot’ from a single pooled moment → not identified from a pooled moment; the joint age-moment fit gives $167 per square foot [154, 182] on a two-storey basis.
Hold-out: calibration that ‘transfers to a geographic hold-out’ → the hold-out does not establish transfer.
Reliability of judgment attributes: now measured within and across extractors (test–retest κ 0.77–0.95 against 0.49–0.63 between models) and with paraphrased prompts.
Sampling weights: stored but applied nowhere in the first version → every sample-level statistic reported unweighted and design-weighted side by side (basement shown in 3.1% of listings, 2.1% weighted; ratio 0.75 → 0.95, 0.70 → 0.91 weighted).
Storeys agreement with the roll: κ = 0.51 → κ = 0.52; raw model responses and the internal analysis plan with its list of deviations are now released.

Limitations stated in the paper

The design does not identify whether listing media add valuation information beyond a correction of the provincial age–life schedule; identifying it would need expert labels of condition on the same listings or a depreciation schedule estimated jointly with the extractor scale.
There are no human expert labels: agreement is measured against broker fields, which are claims, and judgment attributes are evaluated for reliability only, not validity; the proposed expert annotation study was not executed.
Latent attributes such as foundation integrity, hidden moisture or structural soundness cannot be inferred from listing media, and the Monte Carlo interval is miscalibrated against its own target.
Cross-city cost levels are not published and the two unit-cost anchors sit on different storey bases; assessor values are dated judgments rather than truth, asking prices are not transactions, and the comparison with recorded sales rests on 146 quickly sold, cheaper, Montréal-heavy units.

Read online (interactive reader)Revised PDF Revised branch on spbgit

Data

Listings matched to the assessment roll

Read-only copies of the Vrai Prix databases taken on 17 October 2026: the provincial assessment roll (3.75 million units, public under CC-BY 4.0) and house listings offered for sale between 12 August and 17 October 2026. From 39,278 house listings a frame of 7,745 listings matched within 15 m to a single-family roll unit is kept, from which a stratified sample of 1,200 is drawn (135 strata of metropolitan area × age band × asking-price quartile); design weights are stored and applied.

Photographs and extractions

14,245 image URLs (the first twelve photographs per listing in the seller's order) read by a small model on all listings, a mid-size model on the first ten photographs of all listings, and the largest model on a nested subsample of 200 listings and on text. Photographs are read by URL; hosts that do not permit automated retrieval are excluded from the frame.

Benchmarks

Broker-entered structured fields and roll attributes for agreement; indexed assessor building values and asking prices for alignment; a supplementary set of 146 frame listings whose roll unit recorded a sale after listing, for a small test against transaction prices.

Cost schedule sources

Two published per-area construction cost anchors for Montréal and Québec in 2026, Statistics Canada's building construction price indexes by metropolitan area (table 18-10-0289-01) and roll land values indexed to the valuation date; every parameter is sourced or declared.

Methodology

Schema-constrained extraction and deterministic aggregation

Per-image structured extraction with an explicit option to declare a feature not visible, aggregation across images, and reconciliation with text, broker fields and the roll in five baselines, from roll-only to multimodal with administrative override.

Transparent cost model

Indication = indexed roll land value plus replacement cost new, less condition-adjusted age–life depreciation; the improvements component is price-free, and the only price that enters is the transaction-based index applied to the roll's land value.

Reliability measurement

Agreement with broker fields and the roll, test–retest runs, prompt paraphrases and three model tiers, confidence calibration and abstention, and the room-by-position structure of what sellers photograph.

Alignment with paired inference and one-constant baselines

Paired differences in deviation from assessor values, asking prices and recorded sales with municipality-cluster bootstrap intervals, unweighted and design-weighted side by side, compared with roll-only schedules that change a single constant and with a permutation of the extracted attributes across listings.

Sensitivity, calibration and uncertainty

Schedule and depreciation sensitivity, a joint fit of level, economic life and condition strength against assessor values on five metropolitan areas with Québec City held out, Monte Carlo propagation and split-conformal intervals.

Reproducibility

The repository contains the configuration files (data rules, sample design, ontology, prompts, models, cost schedule), the pipeline modules and numbered scripts, unit and regression tests, every table and figure as generated, and a machine-readable manifest with the hash of every artefact.

The raw model responses of every extraction batch are released without images or image URLs, together with the token-usage log, because model outputs differ from run to run and across model versions.

All numbers in the text are macros generated from the result files; the analyses run from the deposited responses, the longest in about a minute on a 16-core Apple-silicon machine.

The extraction runs can be reproduced through the provider's batch interface at a cost of about US$127.5 for the main runs and US$31.74 for the quality-control batches.

Not redistributed: the two Vrai Prix databases (the roll is public on Données Québec and the listing data can be requested from the author under stated conditions), the listing-level analysis files and the batch request files, which carry image URLs.

Cite this paper

Boucher, S.-P. (2026). From Listing Images to Property Valuation: A Multimodal AI Framework for Automated Building Characterization and the Cost Approach. UQO Working Paper No. 20, Université du Québec en Outaouais.

@techreport{boucher2026wp20,
  author      = {Simon-Pierre Boucher},
  title       = {From Listing Images to Property Valuation: A Multimodal AI Framework for Automated Building Characterization and the Cost Approach},
  institution = {Université du Québec en Outaouais, Département des sciences administratives},
  number      = {20},
  type        = {UQO Working Paper},
  year        = {2026},
  month       = {nov},
  url         = {https://www.spboucher.ai/research/wp20},
}