What is an AVM: automated property valuation explained
A full guide: how the model works, what data it needs, how its accuracy is honestly measured, and the cases where it does not replace a human appraiser.
Definition
An AVM — an automated valuation model — is a statistical model that estimates the market value of a property from its characteristics and from market data, without an appraiser visiting the site. It takes an address or coordinates plus the property's parameters, and returns an estimated value, an uncertainty interval and a confidence level.
The key word in that definition is "model". An AVM does not look up an identical sold property in a database: as a rule no such property exists. It estimates what each characteristic is worth separately, and assembles the value from those parts.
What an honest AVM response contains
- A point estimate — a single number, the median estimated value
- A confidence interval — the range the value falls into at a stated probability
- A confidence level — how dense the data is around this particular property
- An explanation — which characteristics moved the result most
A valuation with no interval and no explanation is useless to a bank: it cannot be defended to a credit committee, and it cannot be challenged.
How an AVM works
Four steps, the same in any honest implementation.
Collect and clean the data
Market observations — transactions where they are published, listings where they are not. Deduplication, outlier removal, scoring the quality of each observation.
Enrich with location
Each property is joined to the characteristics of its place: access, infrastructure, environment, risk. Without this the model cannot tell two identical buildings in different districts apart.
Train the model
Hedonic regression, gradient boosting, or a combination. The model learns what a square metre, a floor, a finish and a district are each worth on their own.
Calibrate and retrain
Regular validation on held-out data and retraining as observations accumulate. A model that is not retrained goes stale along with the market.
What data an AVM needs
The main question about any AVM is not which algorithm it uses, but what it learns from.
In countries with an open transaction registry, an AVM learns from transactions. In Uzbekistan transaction prices are not published by law: there is no MLS and no public index built on actual sales. So here the model is built on a panel of listings — and all the difficulty moves into data cleaning, where you must separate the asking price from the achievable one, remove duplicates, and account for the fact that the composition of what is listed changes faster than prices do.
Hedonic regression and machine learning
These are not alternatives but two tools with different strengths. Serious implementations use both.
Hedonic regression
Where it is strong
Produces an interpretable coefficient per characteristic: "a ground floor is worth this many percent less, all else equal". That is what you can show a regulator.
What it cannot do
Handles non-linearities and interactions between features poorly
Gradient boosting
Where it is strong
Noticeably more accurate on the same data, and finds feature interactions by itself
What it cannot do
Gives no ready-made "what a floor is worth" coefficient; explainability has to be built separately
The hedonic approach is not exotic here: the Central Bank of Uzbekistan uses hedonic regression in its own quarterly reports. That means a model built the same way has an external reference point — and any divergence from it can and should be explained in public.
How AVM accuracy is measured
The phrase "95% accurate" means nothing on its own until you know what was measured and on which sample. These are the metrics used professionally.
MAPEMean absolute percentage error. The most common summary metric.MdAPEMedian absolute percentage error. More robust to individual outliers than the mean.PPE10 / PPE20The share of estimates within 10% and 20% of the actual. Answers "how often is the model wrong by a tolerable amount".CODCoefficient of dispersion — a standard mass-appraisal metric: how scattered the estimate-to-price ratios are.PRDPrice-related differential — whether there is systematic bias between expensive and cheap properties.Interval coverageHow often the actual value really lands inside the stated confidence interval. If 90% is claimed and 70% lands, the interval was drawn rather than computed.A single figure is insufficient for another reason: accuracy is not uniform. It is higher where data is dense — the centre of a large city with uniform building stock — and lower on unusual properties and in districts with thin data. An honest model publishes error broken down, not as one number. Our models run at a MAPE of 7–14% depending on the district, on 2025 data.
The trap almost everyone falls into
The most common error in market analysis is not in the model but in how change over time is computed. A raw median price and a composition-adjusted index can diverge as far as the sign.
The reason is that the composition of what gets listed changes faster than the value of the properties themselves. If more large new-build apartments come to market this month, the median rises even if nothing became more expensive. And the reverse. On our Tashkent data in March 2026 the raw median showed a decline while the fixed-weight index showed growth; in the car market a raw year-on-year comparison showed growth while a like-for-like one showed a decline.
The practical takeaway: if someone shows you price dynamics without stating the method, ask whether composition was held fixed. It is the only way to tell a change in prices from a change in the mix.
AVM and appraiser — where the line is
An AVM does not replace an appraiser in every case, and an honest vendor says so plainly.
What is specific to the Uzbekistan market
Three circumstances that make an AVM here different from one built for the US or Europe.
Transaction prices are closed
Prices are not published by law and there is no MLS. The model is built on a listings panel rather than a transaction registry — and it stands or falls on the quality of data cleaning.
Regulatory events, not a smooth trend
Mandatory escrow from April 2026 is an example of a structural break that must not be averaged together with ordinary market movement. Dates like that need to be marked in the model explicitly.
There is an external reference point
The Central Bank publishes its own hedonic coefficients. That is a rare opportunity: a model can and should be reconciled against the regulator's, with the divergences explained.
Frequently asked questions
How does an AVM differ from a calculator on a listings site?
A calculator usually takes an average price per square metre for the district and multiplies it by area. An AVM estimates the contribution of each characteristic separately, accounts for location and returns an uncertainty interval. The difference shows up wherever a property differs from the average — which is most of them.
How accurate is an AVM?
It depends on how dense the data is around the property, and an honest answer is always broken down rather than given as one number. Our models run at a MAPE of 7–14% by district on 2025 data. The question worth asking a vendor is a different one: what share of estimates falls within 10%, and does the actual value land inside the stated interval as often as claimed?
Can an AVM be used for bank collateral?
For pre-filtering applications and for monitoring an already-issued portfolio — yes, and that is usually how it gets deployed. For a final, legally binding valuation the requirements are set by the regulator and the bank's own policy. The practical pattern is that the AVM clears the obvious cases in both directions and the appraiser handles the rest.
What is a confidence interval and why does it matter?
It is the range the true value falls into at a stated probability. It exists to distinguish a confident estimate from an unconfident one: two identical point estimates carrying intervals of ±5% and ±25% represent completely different decision quality. A valuation without an interval cannot be used in a risk process.
How does an AVM account for the district and the location?
Through a location index that turns a coordinate into a set of measurable characteristics of the place — access, infrastructure, environment, risk. Ours is GeoScore, and it is fed into the valuation model as a feature. One important constraint: such an index must not contain price, or the model starts explaining price with price.
Want to test the model on your own data?
We will send the methodology, accuracy metrics broken down by district, and test access to the API