Median Error Rate: How to Read an AVM's Accuracy Claim

A median error rate is the middle value of all the absolute percentage errors a model made: half its estimates were closer to the eventual sale price, half were further off. It is the most common way AVM accuracy is reported, and on its own it tells you nothing about the half of the distribution where money is actually lost. That is not an argument against the metric. Median error is the right headline statistic for this problem, for reasons worth understanding. It is an argument for knowing precisely what it does and does not describe before you make a decision based on it.

How is median error rate calculated?

For every property in the test sample, take the model's estimate and the actual sale price, then compute the absolute percentage error:

|estimate - sale price| / sale price

Do that for every property, sort the resulting percentages from smallest to largest, and take the middle one. That is the median absolute percentage error, usually abbreviated MdAPE or reported simply as "median error."

If a model reports a median error of 6%, it is saying: on this sample, half of our estimates were within 6% of the sale price. It is not saying estimates are usually off by about 6%, and it is definitely not saying the worst case is 6%.

Why median instead of average?

Because the average is dominated by disasters. Home value error distributions have long right tails: most estimates are decent, a handful are catastrophic, and a single estimate that missed by 300% on a mispriced record can move a mean error figure noticeably. The median is robust to those outliers and therefore describes typical performance more faithfully. The trade is that robustness cuts both ways. The same insensitivity that keeps a few bad records from distorting the number also keeps the median from telling you the bad records exist. A model can improve its median while its tail gets worse, and the headline will look like progress.

What does the median hide?

Everything above the fiftieth percentile. Consider two models that both report a 6% median error:

These are illustrative figures, not measurements of any real product. They exist to make one point: the two models are identical on the headline metric and very different to live with. Model B is slightly better at the center and much worse in the tail. If you are screening a hundred properties, Model B is fine. If you are making an offer on one, Model B can ruin a deal in a way Model A cannot. This is why you should ask for hit rates alongside the median. The share of estimates within 5%, 10%, and 20% of the sale price describes the shape of the distribution, and the shape is what determines your realistic downside.

The four questions any accuracy claim must answer

A median error rate with no context is a marketing asset. Four questions turn it into information.

What was the sample?

How many properties, over what period, in which markets, and selected how? Every filter improves the number: on-market only, single family only, complete data only, normal price range only, homes with a recent prior sale only. Each is individually defensible. Stacked, they can move a figure dramatically without any change to the model.

Was the estimate made before or after the property listed?

This is the one that matters most. A model permitted to update after listing is partly reading the answer key. Accuracy measured on refreshed, post-listing estimates and accuracy measured on estimates locked before any market feedback are different metrics that share a name. Published error rates vary by market and are typically far worse for off-market homes than for listed ones, and the difference is usually larger than the difference between vendors.

What was it graded against?

Only one answer is fully credible: the actual recorded closing price.

Is the sample the one you would have chosen in advance?

If a provider selects which properties to report on after seeing the results, the figure is not an accuracy measurement, it is a curation exercise. The defensible version is a rule fixed in advance, applied to everything that meets it. None of them is a lie. This is why "median error rate" as a standalone comparison between vendors is close to meaningless unless the measurement protocol matches.

What about confidence scores and FSD?

Forecast standard deviation, or FSD, is the per-property version of the same idea: the model's own estimate of how uncertain it is about this specific house, expressed as a percentage. It reflects comp density, data completeness, and how far the subject sits from the properties the model knows well. FSD is more useful to you than any aggregate figure, because it is about your property rather than the average property. If a model reports high uncertainty, that is not the model failing, that is the model doing its job. Treat it as an instruction to get a human involved.

How to read our accuracy page

Every Resideline estimate is frozen at the moment a property lists, before any market feedback, and then graded against the real closing price when the sale records. The estimate cannot be revised after the fact, and the sample is not curated after the results are known. We publish the running results at /accuracy rather than quoting a figure in an article, because a number in an article is a snapshot with no provenance and the page is the live scoreboard. It includes the segments where the model performs worst, which is the part of an accuracy report that is actually load bearing. For the mechanisms behind the errors, see why estimates get it wrong and why estimates break on fixer-uppers. To audit an individual number rather than an aggregate, the comps in the CMA report are visible and adjustable.