How Accurate Are Online Home Value Estimates?

Online home value estimates are accurate enough to browse with and rarely accurate enough to transact on. For an ordinary three bedroom house in a dense suburb that is currently listed and sits near the middle of its price range, the major consumer models usually land within a few percent of the eventual closing price; for an off-market house in a thin rural market, an unusual property, or anything that needs work, the same models can miss by a fifth of the value or more. That gap between the good case and the bad case is the entire story, and it is the part a headline accuracy figure flattens. This article covers what accuracy means technically, why published figures are usually measured under favorable conditions, and how to judge one specific estimate instead of trusting a brand.

What does "accurate" actually mean for a home value estimate?

Accuracy for an automated valuation model, or AVM, is normally reported two ways. The first is a median error rate, often written as median absolute percentage error. Take every estimate the model made, compare each to what the home actually sold for, express the miss as a percentage of the sale price, and report the middle value. Half the estimates were closer than that, half were further off. The second is a set of hit rates: the share of estimates that landed within 5%, within 10%, and within 20% of the sale price. Hit rates are more honest than a single median because they describe the shape of the error distribution rather than one point on it. A model with a respectable median can still put roughly one estimate in eight more than 20% off, and if your house is one of those, the median did not help you. A median describes the middle of the distribution, and your house is not the middle of the distribution. There is a longer treatment in median error rate explained.

Why are on-market and off-market accuracy completely different numbers?

This is the most important and least explained fact about published AVM accuracy. Any model allowed to refresh after the listing appears is partly reading over the seller's shoulder. These figures look good. These figures are meaningfully worse, and they are the ones that matter if you are valuing a home nobody is selling yet. Published error rates vary a great deal by market, and off-market figures are typically far worse than on-market figures. When a provider reports one number without saying which question it answers, assume the more flattering reading. This is why we freeze. A Resideline estimate is locked at the moment a property lists, before the market has voted, and then graded against the actual closing price when the home sells. A frozen estimate cannot quietly drift toward the answer. The running results are published on the accuracy page rather than described in prose here.

When are online home value estimates most reliable?

There is a hard ceiling no amount of modeling removes. A system working from public records and listing data cannot see:

• •Condition. Public data records that a house exists, its size, its bed and bath counts, and its year built. It does not record that the roof is at the end of its life or that the kitchen was gutted six years ago. - Interior quality. Two identical floor plans on the same street can be far apart on finish level alone, and the assessor's file says they are the same house. - Micro location. Backing onto a highway, a retention pond, or a commercial lot. Some of that is inferable from coordinates. Much of it is not. - Terms and motivation. A cash close in nine days, a seller credit, an estate that needed to move by a date. An honest AVM narrows the range on what it can see and states plainly where its assumptions live. That is the reasoning behind two of our design choices: comps are visible and adjustable rather than hidden inside a score, and the condition read comes from listing photos, labeled as an estimate rather than an inspection.

How do I tell whether a specific estimate is trustworthy?

Do not check it against another estimate. Two AVMs agreeing tells you they picked similar comps, not that either is correct. Instead:

1. Open the comps. If you cannot see which sales produced the number, you cannot audit it, and you are trusting a brand rather than a method. 2. Check that the comps are comparable. Similar living area, age, property type, and school and tax boundaries. A common working band is plus or minus 20% of the subject's square footage. 3. Check the geography. A sale half a mile away can be in a different submarket if a highway, a district line, a flood zone, or a different HOA sits between you. 4. Check the condition assumption. If the model assumed average condition and the house is not average, the estimate is wrong in a direction you can predict. 5. Check the dates. Comps older than about six months in a moving market describe a market that no longer exists. The full walkthrough lives in how to check whether a home value estimate is accurate.

So what number should I actually use?

Use a range, not a point. Any single figure implies a precision the underlying data cannot support. In practice:

• •Browsing or screening. An AVM is excellent. You are sorting, not deciding. - Underwriting an investment. The estimate belongs inside a model, not inside a decision. Pair it with a rehab budget and run the deal, which is what the deal analyzer and the free ARV calculator exist to do. - Lending, tax appeal, or litigation. You need an appraisal. An AVM is not a substitute and no serious vendor claims otherwise.

The honest summary

An online home value estimate is a fast first draft written by a system that has never seen the house. On ordinary homes in busy markets with plenty of recent sales, these models are genuinely close. On the properties where most of the money is made and lost, distressed, unusual, rural, or mid renovation, they are much weaker, and the weakness runs in a predictable direction. What is worth demanding is not a better number but a visible track record. Ask any provider when the estimate was made, what it was graded against, and whether you can see the sample. Ours is published at /accuracy and updates as homes close. The underlying tools run free without an account from free tools.