Industry commentary puts it plainly: “AI-driven Automated Valuation Models, or AVMs, are transforming how we analyze property markets and determine values.” What the commentary rarely separates is the two things actually being bought — the valuation model, and the data feeding it. They fail differently and they are tested differently, and conflating the two is how weak AVMs get purchased.
For clarity about this site's own position: BrightCat supplies Canadian property data, updated weekly. It does not produce valuations. A bank, insurer or proptech building or buying an AVM uses data like this as input and reference. The model, and the accuracy claim, belong to whoever built it.
A Canadian AVM draws on five kinds of input: property attributes, which carry the hedonic backbone; listing activity, which is the market's live pricing experiment on each property; confirmed transactions, which are the ground truth the model is ultimately judged against; rental signals, for yield-based checks on investor stock; and a market index, to carry past sales forward through time without composition distortion. The first question for any model is which of these it actually has, at what freshness, for the geography it claims.
The data side gets the same checks as any model input. Cadence: BrightCat captures weekly, typically within 7 days of listing activity. Identifier persistence: a property that relists under a new MLS number must remain one history, not several fragments. Coverage: confirmed sold prices exist in Ontario, British Columbia and Alberta, which also defines where ground-truth validation is possible. The repeat-sale layer holds 194,167 verified pairs across 12 years, the raw material of a repeat-sale index or a time-adjustment step. And the output side is a plain engineering question: a valuation with a confidence range per property, refreshed on a stated schedule, is a feature a portfolio model can consume and backtest.
Comparing the accuracy of Canadian AVMs only means something when every model is measured the same way on the same held-out sales. Published accuracy figures measured on different samples cannot be ranked against each other, which is why this page ranks nobody.
A fair test covers five things. Out-of-sample error on sales the model never saw, usually median absolute percentage error plus the share of valuations within 10% of the sold price. Error sliced by province, city, property type and price band, because a model can be strong in Toronto condos and weak everywhere else while the headline number hides it. The age of the inputs on valuation day, and what that lag costs in a moving market. Whether the model says when it does not know, and whether its stated confidence matches realised error. And backtests through rising and falling markets, not just the last flat quarter.
Confirmed sold records are the natural hold-out for that test, and a repeat-sale index is the reference series for time adjustment: a model whose valuations drift against a composition-controlled index is drifting. Reported dissatisfaction with any single valuation tool fits the same frame — one user's experience is evidence about one workflow, and a reason to examine limitations rather than a verdict on the market. Anyone evaluating an AVM can request the same sample data a model builder would use and run the hold-out test themselves.
Real Canadian property data, delivered through Snowflake, MCP, or flat file.