What we show
Condition-aware market value estimate, range, and seller-ask gap for the selected vehicle.
Methodology
How the estimate is built and tested.
Every figure here is documented: what it measures, how much evidence supports it, and where it can be wrong.
Claim boundary
What condition-aware market value can current Canadian evidence support?
The estimate is a current asking-market anchor adjusted for relative condition and odometer. It does not guarantee the completed transaction price.
What we do not show
Known transaction value, certified appraisal, recommended offer, or residual forecast.
Why the range
The range combines the asking-market spread in the Canadian anchor with condition and odometer effects from later-year models.
Minimum evidence
A price cell is shown only when the Canadian anchor has n ≥ 10 dealer listings.
Condition evidence
Six condition selections form one auction-grade equivalent. The model learns one effect for that composite grade instead of six separate dollar adjustments.
Current Canadian anchor
5,605 province × make × model × year cells from dealer asking prices.
SOURCE: OMNIAAUTO CANADIAN AGGREGATESCondition + odometer model
Trained on 91,278 completed US wholesale outcomes with close-peer matching and a gradient-boosted residual.
SOURCE: LARSEN NBER AUCTION OUTCOMESDisclosed range
The range combines the Canadian source spread with temporal-test residuals. It does not guarantee coverage of Canadian sale prices.
SCOPE: ASKING-MARKET ESTIMATE ONLYValidation (temporal-test)
Train2006–2008
Test2009–2010
Held-out outcomes39,132
MAE$1,198
WAPE11.60%
vs peer baseline4.29% better
CONDITION MODEL · USED IN RESULTS
How the condition model is trained and tested.
Outcome and target
The model compares 91,278 completed US wholesale outcomes with leave-one-out peers matched on sale year, auction, vehicle year, make, model, and trim code. It predicts log sold price relative to that peer anchor.
Condition input
The overall grade, accident/title, mechanical, cosmetic, service, and wear selections form one bounded auction-grade equivalent. Gradient-boosted trees learn how that composite grade and the odometer reading affect price. Average grade at the anchor odometer is neutral.
Temporal test
Training uses 2006-2008 and testing uses 39,132 outcomes from 2009-2010. With the deployed centering, MAE is $1,198 and WAPE is 11.60%. MAE improves 4.29% over the peer-only baseline. These are historical US wholesale metrics and do not measure Canadian retail accuracy.
Inspect the model metrics →DATA PROVENANCE
Source, licence, and build-time checks.
Market source
OmniaAuto’s Canadian Vehicle Market Aggregates describe 624,678 vehicles across Canada. The data is aggregated to protect privacy and keep statistics stable. Price statistics come from dealer asking prices, and duplicate vehicles are resolved to an originating seller.
View source dataset ↗Licence
The aggregate dataset is released under CC BY-NC 4.0 for research, teaching, and non-commercial use. AutoValue Canada is a non-commercial portfolio demonstration and credits the publisher on every surface where the data appears.
Read the licence ↗Build-time QA
The pipeline verifies schema, numeric types, unique market-cell keys, minimum sample size, positive prices, and monotonic percentile order. A failed check stops the build.
Inspect the QA summary →MODEL CARD · RESEARCH ONLY
How the research benchmark is validated.
Target
The target is the published median dealer asking price for an aggregate province × make × model × year cell. The benchmark is evaluated separately from the consumer lookup and never changes a displayed source statistic.
Leakage control
Five-fold GroupKFold validation holds out complete make-model groups. No make-model pair appears in both training and evaluation data within a fold. All reported scores are out-of-fold.
Result and limit
Histogram gradient boosting cuts weighted MAE by 45.2% against a training-fold global-median baseline. Its $8,026 MAE is still too high for individual appraisals, and this gap stays visible by design.
View metrics and fold errors →Roadmap (data-gated)
Blocked features and next steps.
BLOCKED
Transaction-value estimate
Needs Canadian completed-sale data.
BLOCKED
1–3 year residual forecast
Needs longitudinal outcome panels.
NEXT
Safety recall context
Transport Canada first-party recall data.
NEXT
Operating-cost context
NRCan fuel and EV ratings.