How MdAE is calculated
For every property in a benchmark set of recent sales:
- Run the AVM on the property as if you didn't know it had sold.
- Take the percentage error:
|estimate – sale_price| / sale_price. - Collect that absolute percentage error for every property.
- Take the median of the resulting list.
That's MdAE. A 5% MdAE means half the AVM's estimates were within 5% of the true sale price, half were further off.
Why median, not mean
Sale prices have fat tails. A handful of trophy homes that an AVM gets badly wrong (a $40M Malibu beachfront the model has no comps for, a 200-acre ranch with no comparable inventory) can move the mean absolute error by several percentage points. The median ignores those outliers and tells you what accuracy a typical buyer should expect on a typical home.
The same reason real-estate professionals quote median sale price in a market report rather than the mean — a few mansions distort the mean and don't represent the buyer at the median.
Vendors that quote mean absolute error are usually doing one of two things: showing a number that flatters their model on a benign benchmark set, or quoting active-listing-only data. Neither is what you actually care about as a buyer.
What counts as good MdAE
| MdAE band | Interpretation |
|---|---|
| < 3% | Lender-grade. Quantarium, HouseCanary on dense markets. |
| 3–5% | Lender-acceptable. Most large vendors here. |
| 5–8% | Consumer-grade. Zestimate, Redfin Estimate, Twellie off-market. |
| 8–12% | Use the confidence band, not the point estimate. |
| > 12% | The model is guessing. |
A formal appraisal — the legal benchmark — clocks in at 1–3%. That gap between an AVM and an appraisal is real, and it's the reason lenders fund off appraisals.
The active-listing trap
Almost every consumer-facing AVM publishes two MdAE numbers, and the prettier one is for active listings — homes currently on the market with photos, descriptions, and recent comp signal. The harder benchmark is off-market — a home that hasn't sold in 5–10 years and has no current photos. Off-market is the number that matters when you're shopping, because that's the population you're shopping.
Zillow's frequently-cited 1.9% Zestimate accuracy is the active number. The off-market number is closer to 7.5%. Both are "true"; only one is relevant to a buyer comparing off-market homes in a neighborhood.
How MdAE relates to confidence bands
MdAE is a population-level metric — accuracy across thousands of homes. The confidence band on a single AVM estimate is property- level — how confident is the model on this one address.
The two are correlated but not identical. A vendor with 6% MdAE will publish wider confidence bands on atypical homes (custom builds, thin-comp markets, stale public records) and tighter bands in tract subdivisions where every home is a near-twin. Read both. A 4% MdAE with a $50k confidence band on your target home means the neighborhood is hard, even if the vendor is generally accurate.
For a deeper read on what to do with the confidence band see How to find and read comparable sales.
Watch for these dishonest framings
- "94% accurate." This is 1 minus the percentage of estimates outside a 5% (or 10%) band. Looks great, hides the median.
- Reporting median percentage error, not absolute error. Equal-but-opposite errors cancel. A model that's +20% on half the homes and -20% on the other half has 0% median percentage error — and is useless.
- Cherry-picked benchmark sets. Active-listing-only, single- metro, or dense-tract-subdivision benchmarks all flatter the number. Ask what population the MdAE was calculated on.
- No band. Any AVM that quotes a point estimate without a confidence band is failing the buyer. The band is what makes the number actionable.
The honest version of this is: pull a report, read the MdAE for the relevant population, then look at the confidence band on your specific home. If the vendor only gives you the headline, check their methodology page or treat the number as informal.