How Accurate Is AI Property Valuation? Median Error, PPE10 and Hit Rate Explained

Get a summary of this article with your favorite AI
AI Summary
- •AI property valuation accuracy is judged by tests, not claims: median error, PPE10, hit rate, forecast standard deviation (FSD) and bias, each split by geography, price tier and property type.
- •Zillow's published nationwide off-market median error is 7.21% (October 28, 2026 refresh), a consumer benchmark rather than a professional standard.
- •To test any tool, run a blind back-test on recent arm's-length sales and check whether its confidence scores hold up.
- •Set a tolerance for each decision first, and send values outside it to human review.
How accurate is AI property valuation? The most visible public number comes from a consumer portal: Zillow reports a nationwide median error of 7.21% for homes that aren't listed for sale (refreshed October 28, 2026), and its page doesn't spell out the benchmark behind that figure. On a $500,000 home, that would mean half of off-market estimates land within roughly $36,000 of the value they're measured against, and half miss by more (illustrative math).
That's a fair yardstick for a free estimate. Professionals need proof that AI property valuation accuracy holds on their own markets. Below: how to measure property valuation accuracy, which metrics to ask for, and how to test any tool on your own closed sales.
How accurate are AI property valuations for residential homes?
It depends on the property, the data and the market, and no industry-wide figure exists for AI real estate valuation accuracy. Estimates tend to hold up where recent sales are plentiful and records are complete, and weaken on unusual homes, thin markets and stale data. Zillow's own page says accuracy varies by location and property because it depends on the depth and quality of available data. The wider workflow is covered in the AI property valuation guide.
What is a good accuracy rate for an AVM? There is no universal accuracy threshold for an automated valuation model (AVM). The acceptable error depends on the intended use, local market, property type and cost of a wrong valuation. Test on a representative set of closed residential sales and set your acceptance threshold before testing.
How can AI enhance property valuation accuracy? It can help by drawing on more records, recent sales and market trends than a manual comp search usually covers, and by flagging data gaps. Whether that improves the accuracy of AI property valuation models for you shows up only in testing.
Is an AI property valuation the same as an appraisal?
No. An automated estimate isn't a licensed appraisal or a defensible professional valuation by default. Zillow states that the Zestimate "is not an appraisal and can't be used in place of an appraisal," and that most lending professionals use only professional appraisals for loan decisions. AI home value estimate accuracy depends on four things:
- Data. Zillow names county and tax assessor records and MLS feeds among its inputs. Ask any tool what it draws on.
- Local market. AVM accuracy by location varies, so a national median says little about a rural county.
- Property characteristics. Unusual homes have fewer true comparables.
- Intended use. A listing-price guide and a loan decision need different proof.
Which accuracy metrics should you ask a vendor for?
To measure property valuation accuracy, ask for six: median absolute error, the share within ±10% (PPE10), mean absolute error, signed bias, hit rate or coverage, and a calibrated confidence indicator such as FSD. These property valuation model accuracy metrics answer different questions, so they aren't interchangeable. Request each split by location, price tier and property type.
Convention here: error = (value − benchmark) ÷ benchmark, with the recorded sale price as the benchmark; "absolute" drops the sign.
| Metric | Definition | What it can hide | Ask the vendor |
|---|---|---|---|
Median absolute error (AVM median absolute error) | Median of the absolute errors: half of values miss by less, half by more | Size of the worst misses | What is the mean absolute error alongside it? |
PPE10 | Percentage of properties whose estimated value lands within ±10% of the benchmark | How far off the rest are | What are PPE5 and PPE20? |
Mean absolute error | Average of the absolute errors | One big miss pulls it up | Is it reported with the median? |
Signed bias | Median of the signed errors: positive means values run high, negative means low | Areas that run high offsetting areas that run low | What is it by neighborhood and price tier? |
Hit rate or coverage | Share of requested properties that return a value | A tool can look accurate by skipping hard properties | How is it counted, and which inputs are required? |
FSD or calibrated confidence indicator | The model's own forecast of how far a value could miss; "calibrated" means checked against sales | No standard FSD scale across vendors | Has AVM confidence score accuracy been checked on closed sales? |
Definitions vary by source, and no US rule fixes one. When federal agencies finalized their AVM quality control standards, effective October 1, 2025, commenters asked for standardized confidence scores and testing standards, and the agencies declined to adopt prescriptive requirements. So ask every vendor how each metric is defined.
What is the difference between median error and PPE10?
Median absolute error measures how big the typical miss is. PPE10 measures how often a value lands within ±10% of the benchmark. Two tools can tie on one and differ widely on the other, so read both with mean absolute error, bias and the largest misses.
So, what percentage of AI property valuations should fall within 10% of the actual sale price? No universal share exists. Set a target for each decision before testing (see the tolerance table below), and read PPE10 only alongside the sample it came from.
Illustrative example, with hypothetical numbers from no real tool: 10 closed sales, each recorded sale price as the benchmark, and both tools valuing every home as of a date before the sale without seeing the sale price.
| Metric | Tool A | Tool B |
|---|---|---|
Median absolute error | 6.0% | 6.0% |
PPE10 | 100% | 80% |
Mean absolute error | 5.9% | 9.1% |
Largest single miss | 9% | 35% |
Absolute errors: Tool A, 2, 3, 4, 5, 6, 6, 7, 8, 9 and 9%. Tool B, 1, 1, 2, 3, 6, 6, 7, 8, 22 and 35%. Both show the same median property valuation error rate, yet Tool B missed two homes by 22% and 35%. On a $400,000 home, a 35% miss is $140,000.
Plan Your Valuation Workflow
Walk through your data sources, review rules and integrations with the Svermo team.
How can a real estate professional test an AI valuation model using closed sales?

Run AI property valuation back-testing: take recent arm's-length sales, have the tool value each property as of a date before the sale without seeing the price, compare each value to the actual price, and report the metrics by segment. This is also how to validate an automated valuation model before relying on it.
- Build a representative, recent sample. Use arm's-length sales only, meaning no family transfers or discounted deals. The IAAO standard on AVMs says test samples should approximate the property population and that larger samples give greater precision.
- Keep it blind. Ask for values dated before closing. The IAAO standard calls for testing on sales not used in calibration, known as a holdout sample. In a 2005 article, Freddie Mac's collateral modeling director suggested sales so recent they haven't reached public records, such as loans closed in the last 30 days.
- Calculate every error and the six metrics, the same way for each tool.
- Check the confidence claims. Per the same Freddie Mac article, an FSD of 10% implies about 68% of sale prices within ±10% of values, about 80% within ±13% (1.28 FSDs) and about 90% within ±16.5% (1.65 FSDs). A lower share means the score overstates precision.
- Split the results. Break them out by location, price tier and property type, and run an AI property valuation bias assessment by comparing signed error across neighborhoods (see testing for disparate error).
How to validate AI property valuation accuracy: checklist
- Select representative, recent residential closed sales.
- Exclude sales used to calibrate the model.
- Request valuations dated before each sale.
- Compare estimates with the actual sale prices.
- Calculate the metrics the same way for every tool.
- Break results down by market and property type.
- Define in advance when a human reviewer must step in.
Why do AI property value estimates differ between tools, and how do you compare them?
Tools differ because each AVM has different performance characteristics, geographic coverage and confidence scores, and published numbers differ because each reflects a specific test. To compare automated property valuations fairly, and to compare two tools head to head, hold four things equal:
- The sales. Same closed sales, same blind conditions.
- The benchmark. Sale price, appraised value and listing price give different answers.
- The metrics. Hit rate, coverage and success rate aren't the same measure, so accuracy vs hit rate only compares with matching definitions.
- The confidence scale. Some AVMs don't tie their confidence scores to FSD at all, per Freddie Mac's modeling team.
How much accuracy is enough for your use case?
Set a tolerance for each of the use cases of AI real estate valuation software before testing, based on what a wrong value costs.
| Decision | What a miss costs | Set before testing |
|---|---|---|
Listing price | A listing that sits, or money left behind | A defensible pricing band; wide-variance homes go to an agent |
Loan or collateral decision | A loan above the property's real value | A maximum error, a minimum confidence level and a review rule below it |
Portfolio revaluation | Value drift hides risk | Sample validation on a schedule and bias tracking by segment |
So, when should a property professional request a human review of an AI-generated valuation? When a value falls outside the tolerance you set, when the confidence indicator flags data gaps or high variance, when the property is unusual or the data is thin, and when the value will feed a lending or legal decision where a licensed appraisal may be required. Seven situations where an AI value needs human review covers the usual triggers.
For banks, federal regulators replaced SR 11-7 with SR 26-2 on April 17, 2026. The guidance scales testing to a model's materiality and says model risk principles still apply when a bank doesn't receive a vendor's code, data or methodology.
What should an accuracy report from your valuation tool show?
At minimum: a value, a range around it, the evidence behind it, a confidence indicator checked against closed sales, segment-level results, data freshness, and a record of who reviewed or changed it.
Valuation AI lists comparable property analysis, confidence insights that highlight data gaps and valuation variance, valuation audit logs, and workflows that can be adapted to an organization's review and approval process. The product page shows them as features and claims no accuracy figures, so test any tool, this one included, on your own sales for an accurate analysis.
Check the Number Behind Every Value
See how Valuation AI shows the comparable sales, confidence insights and review trail behind each property value.
Ready to Let Your Valuation Tool Face the Closed Sales?
Vendor accuracy claims are a lot like fishing stories: the one that got away is always the biggest. Closed sales don't exaggerate. Hide the prices, let any tool take its shot, and you'll hold something no brochure can hand you: a number you ran yourself.
Want a second pair of eyes on your test file? Reach out to the Svermo team.
FAQs
PPE10 is the percentage of valuations within ±10% of the benchmark, usually the sale price. A PPE10 of 70% means seven in ten values were that close. It doesn't show how far off the other three were.
Hit rate is the share of valuation requests that return a value. Read it with accuracy, because a tool can raise one by lowering the other.
FSD is a model's own forecast of how far a value could miss the sale price. With an FSD of 10%, about 68% of sale prices should land within ±10% of the value.
Zestimate accuracy for off-market homes is a nationwide median error of 7.21%, according to Zillow, as of its October 28, 2026 refresh. Zillow adds that a Zestimate is not an appraisal.
Often as an input, rarely as the whole decision. The reliability you need depends on the stakes: screening many listings tolerates wider misses than setting an acquisition price. Check tested error on similar assets, the worst misses, and whether the tool flags thin data. An accuracy score alone doesn't guarantee a sound investment or pricing decision.
