Automated valuation models: what is behind an online price estimate

In this article

Written by

Kooky

Builder of Shaka, the payment router that pays every agent their commission on closing date.

About Kooky and Shaka →

Type almost any Queensland address into a property site and a dollar figure appears beside it, often with a range and a small label such as high or medium. Nobody visited the house to produce it. The figure comes from an automated valuation model, usually shortened to AVM: software that estimates what a property would sell for today from the records it holds about that property and the ones around it.

These estimates now shape real decisions. Owners use them to decide whether to sell, buyers to judge an asking price, and lenders to check the security for a loan. Yet the same home can show two quite different figures on two sites, and the label that comes with each is rarely explained. This guide covers how an estimate is produced, what data goes into it, how the providers themselves describe its accuracy, why the results differ, how the banking regulator expects lenders to treat it beside a desktop assessment and a full valuation, and why it is weakest on an unusual home or in a quiet market. It does not go back over suburb medians or price indexes, which have their own guides.

First two figures from PropTrack's guide "Demystifying the AVM" (undated); the third from the Cotality (CoreLogic) Australian help centre page on its Confidence Indicator. FSD is the forecast standard deviation, explained below.

What an automated valuation model is

PropTrack, the data arm of the group that owns realestate.com.au, describes an AVM in its guide "Demystifying the AVM" as an algorithm that instantly produces an estimate of a property's value. The guide says the model is built to mirror the way a human valuer reasons: it assembles comparable properties, looks at the home's own sales history, takes account of what is currently listed and forms a view of the property's condition.

Three features set an AVM apart from a valuer's work. It is immediate: PropTrack says a lender connected to its system receives an answer within a fraction of a second. It is cheap to run at scale, so it can be produced for millions of addresses at once instead of one at a time. And it involves no inspection. The model knows only what has been recorded, and it cannot open the front door.

An AVM is also different from the market appraisal an agent gives a prospective seller. An appraisal is an opinion formed by a person who has usually walked through the property. An AVM figure is a statistical output. It comes with a measure of its own uncertainty, which is the part most readers skip and the part that matters most.

The data that goes in

An estimate can only be as good as the records behind it. Going by the providers' own descriptions, the inputs fall into a few groups.

The first is sales. Domain's help centre, on a page dated 26 May 2026, says its price guide draws on official property sales information from state and territory governments, on Australian real estate agents and on the data business Pricefinder. PropTrack's guide adds a point about timing: it says it uses sales advised by agents through realestate.com.au, so that a sale can reach the model before it appears in official records.

The second is the property's own attributes: land size, bedrooms, bathrooms, parking, the type of dwelling. These let the model compare like with like.

The third is listings. PropTrack says current advertisements on its portal feed its model, because the asking prices and the response to them carry information about today's market that past sales cannot.

The fourth is location. PropTrack's guide refers to geospatial data, meaning the home's position relative to amenities and the standing of the neighbourhood.

The fifth is newer: photographs. PropTrack says it applies computer vision to listing images to score a property's presentation, whether it has been renovated and the condition of its features. That is an attempt to close the largest gap in any desk-based method, which is that it cannot see inside. It works only when photographs exist, and only as well as the photographs reflect the house today.

PropTrack also lists what it calls market behavioural data: signs of supply and demand in an area that influence how sought-after a property is.

How the sub-models are blended

An AVM is not one formula. PropTrack's guide describes a set of sub-models whose answers are merged by an algorithm into a single figure. It names five.

A comparable-based model does what an agent does with recent sales: it selects similar properties nearby and adjusts for the differences. A page on PropTrack's site about its current model, last modified in November 2024, says the comparable approach draws on hundreds of thousands of properties.

An indexation-based model starts from the last price the home itself sold for and moves it forward by the change in the local market since. It is strongest when the home sold recently and nothing has changed.

A listing model reads the asking price and the advertisement when a property is on the market.

A hedonic model prices the attributes. It works out what the market has been paying for an extra bedroom, a larger block or a particular location, and adds those components up for the subject property.

An image-scoring model turns photographs into a rating of condition and appeal.

Each has a blind spot. Comparable sales fail where nothing similar has sold. Indexation fails where the house has been rebuilt since it last changed hands. A hedonic model misses whatever is not captured as an attribute. PropTrack's guide puts the reasoning for combining them plainly: the leading providers, it says, use several models to spread the risk of error and remove bias. The same PropTrack page says the model can be retrained several times in a single day.

How accuracy is expressed

Every estimate is wrong by some amount. The useful question is by how much, and how often. Providers answer it in two ways.

The first is a measure attached to each individual estimate, the forecast standard deviation or FSD. PropTrack's guide says the FSD quantifies the confidence in the estimated value: a lower FSD signals more confidence, a higher one more uncertainty. It gives the rule that goes with it. About 68 per cent of the time, the contract price or a formal valuation should fall within plus or minus one FSD of the AVM figure.

A worked example shows what that means. Take an illustrative estimate of $900,000 with an FSD of 10 per cent. Ten per cent of $900,000 is $90,000. The model is then saying that roughly two times in three the true price should land between $810,000 and $990,000, and that about one time in three it will fall outside that band. Halve the FSD to 5 per cent and the band narrows to $855,000 to $945,000. These numbers are invented for illustration and describe no real property.

The second is a measure of the whole model, taken by comparing past estimates with the prices later achieved. The common form is the share of estimates that fell within 10 per cent of the sale price. PropTrack's guide states the benchmark it considers best in class: more than 80 per cent of estimates within 10 per cent of the contract price. That is a statement about what a leading model should do, not a published result for a particular state or period.

Read the benchmark from the other side and it is a caution. Even at that standard, up to one estimate in five is more than 10 per cent away from the price. On a $900,000 home, 10 per cent is $90,000.

None of the consumer pages read for this guide publishes a median error for Queensland or for any other state. Domain's help page gives accuracy bands in words only, and the Cotality help page gives no numerical thresholds for its levels. Where a provider does publish a median error figure, the things to check are the period it covers, whether homes that were on the market are included, and how many properties were left out because no estimate could be made.

What the confidence labels mean

The public rarely sees an FSD. It sees a label. The scales differ from one provider to the next, and so does the amount of explanation behind them.

Cotality's help centre, which still carries the CoreLogic name, defines its indicator as a statistical calculation of the probability that the estimated value is accurate compared with market value. It adds that the indicator should be considered alongside the estimate, and that more local market data generally brings a higher rating.

The common thread is that a label describes the evidence, not the house. A high rating does not mean the property is desirable or that the figure is right. It means there were plenty of recent, similar sales to work from. A low rating means the model was short of material. Because the scales are not aligned, a high from one provider and a high from another are not the same promise.

Domain's page adds one more outcome: no figure at all. It says a property has no estimate when one of acceptable accuracy cannot be generated because the data is insufficient. A blank is the model declining to guess.

Why two sites give different figures

Once the method is clear, the disagreement between providers stops being surprising. It follows from at least four differences, each visible in the material above.

The data is different. One provider receives agent-advised sales through its own portal; another relies on government sales records, agents and a data supplier. A sale that one model has already counted may not have reached the other.

The attributes are different. If one database records a home as three bedrooms and another as four, the two models are valuing different houses.

The models are different. Each provider weights its sub-models in its own way, and not all of them score photographs or read live listings.

The timing is different. PropTrack describes retraining within a day; Domain says its estimates are updated whenever property data is refreshed, without giving a schedule.

There is also the matter of what is shown. A single figure, a range, and a range with a midpoint are three presentations of the same uncertainty, and they leave different impressions.

PropTrack's own guide treats this spread as normal. It reports that most lenders use more than one AVM provider, and it recommends doing so to diversify risk and to cover more properties. For a household, the equivalent is to treat a gap between two sites as information. A wide gap is itself a sign that the property is hard to estimate.

Estimate, desktop and full valuation

An AVM is the lightest of several ways to put a value on a home. The Australian Property Institute, the professional body for valuers, sets out the difference between the two heavier ones in its Standards Series No. 4, published on 3 November 2025.

That standard describes a full valuation process as one that includes all the necessary enquiries and investigations, including a physical inspection of the property by the valuer responsible. A desktop process is evidence-based but involves no inspection, and results in a report containing what the institute calls an indicative value. The standard notes that such reports go by many names in the market, among them remote valuations, virtual inspection reports, hybrid AVMs and valuer-assisted AVMs.

The institute places conditions on the desktop. Only members holding its valuer certifications may produce one, and the valuer must already know the relevant market from having inspected similar properties before. The report must disclose that the valuer did not visit or inspect the property, must set out the limitations imposed by the instructions, and must make clear that the opinion is indicative only.

The practical difference between an AVM and a desktop is the person. In a desktop, a qualified valuer has looked at the evidence and put a name to the figure. In an AVM, no one has.

What APRA expects of lenders

Lenders are the heaviest users of automated estimates, and their use is supervised by the Australian Prudential Regulation Authority. APRA's guidance sits in its prudential practice guide APG 223 on residential mortgage lending. The version read for this guide is dated February 2017, and the paragraph numbers below are taken from it.

The guide accepts the lighter methods in principle. Paragraph 69 says desktop assessments, kerbside assessments, automated valuation methods and reviews of contracts of sale are all acceptable valuation assessments in the appropriate context. The qualification follows at once: as the risk in the security rises, or as the loan is less well covered by the property, a specialist valuation becomes more necessary. Paragraph 72 calls a full on-site valuation good practice, while recognising that a lender may tailor its policy to circumstances.

Paragraph 70 sets out what a lender should have in place before it relies on a model. In summary:

1. A hierarchy of acceptable valuation methods, matched to the level of risk.
2. An analysis of each model's strengths, weaknesses, methodology and data sources.
3. Back-testing of a statistically random sample of AVM and desktop results.
4. Clarity about what the model's outputs mean and how they fit into the lending process.
5. Ongoing monitoring, with periodic back-testing by the lender itself to validate reliability independently.
6. Training for the staff who use the tools.

Back-testing simply means checking past estimates against what the properties later sold or were valued for. APRA expects the lender to do that for itself, not to take the provider's word.

The same guide deals with independence. Paragraph 71 says a lender's panel of valuers should be reviewed periodically by senior risk management staff and that valuers should be selected by the risk area, not by sales staff. Paragraph 73 describes attempts to pressure valuers to over-value properties as an indicator of poor practice.

How much of the work is automated is a matter for each lender. PropTrack's guide says most lenders use AVMs for 30 per cent of their valuation decisions. That is the provider's own figure, given without a date. In practice the idea in paragraph 69 governs: the confidence measure acts as a gate. An estimate with a low FSD on an ordinary home with a comfortable deposit may be accepted; a weaker one, or a riskier loan, is passed up the hierarchy to a desktop or to a valuer who inspects. This is why a borrower may never hear that a valuation happened at all, while another on the same street waits for a valuer's visit.

Unusual homes and thin markets

Every part of the method depends on comparison, so the model is weakest wherever comparison is hard.

Domain's help page says it directly: estimates can be less accurate if the property is unique or does not have many similar properties to compare it to in the area. That covers more of Queensland than it might seem. A house on acreage, a home on a large or irregular block, a heritage building, a one-off architect design or a property in a small town where few homes change hands in a year all leave a comparable-based model with little to hold.

A thin market affects the other sub-models as well. An index built on few sales moves erratically, so rolling an old price forward is less reliable. A hedonic model has too few transactions to learn what local buyers pay for each feature.

Then there is what the records do not show. A model that has no recent photographs cannot know that a kitchen was replaced, that an extension was added, or that a house has been left to run down. Image scoring helps only where a recent listing exists. Nor does an AVM read a building report, inspect for termites or notice the view.

Wrong attributes are the most ordinary cause of a poor estimate, and the hardest for an owner to fix. Domain's page says property attributes can be corrected only by submitting a form, that details an owner enters in its owners' area are for personal reference and do not change the public figure, and that there is currently no way to update a property estimate directly.

The confidence label is the model's own warning about all of this. A low rating, or no estimate, on an unusual home is the system working as intended.

What an estimate can and cannot tell an owner

An online estimate is a fast, free summary of what the recorded evidence suggests. It is a reasonable first look at an ordinary home in a suburb with steady turnover, and a way to see roughly where a property sits. Read with its range and its confidence rating, it is honest about its own limits.

It is not a valuation in the sense valuers and lenders use the word. The Australian Property Institute reserves that for a process that includes inspection, and treats even a valuer's desktop opinion as indicative only. APRA's guide treats the automated figure as one acceptable tool among several, to be tested and monitored, with a full valuation as good practice when the risk calls for it.

Nor is it a prediction of what a particular buyer will pay on a particular day. A sale price is set by the people who turn up. How much weight an estimate deserves in a decision to sell, buy or borrow depends on the property and the circumstances, which is a question for the professionals involved in that transaction.

The most informative part of an online estimate is not the dollar figure but the rating beside it, which says how much evidence the model had.

Kooky, from Shaka

Kooky edits Queensland Estate and builds Shaka, the payment router he made for Queensland property professionals. One payment comes in, and every agent, agency and party in the deal receives their signed share on closing date.

About Kooky and Shaka →