What Is An Automated Valuation Model (AVM)? - HomeLight

How do automated valuation models work?

Generally speaking, there are two ingredients to a value report produced by an AVM: a massive amount of data about as many real estate properties as possible, and a proprietary algorithm based on machine learning and regression analysis.

Types of data used in automated valuation models

In general, AVMs use two datasets to produce the most accurate property estimates: an initial dataset to calibrate the model, and an ongoing, expansive dataset to predict each subsequent property.

The calibration data set includes a representative sample of the types of properties that the model will be estimating and also brings in a reliable source of truth for the valuation, such as recently sold home prices. The source of truth data is instrumental in allowing the model to identify which of all of the other data points available for the property are most effectively combined together in order to make a reliable prediction.

The ongoing dataset pulls in as much information about the property as possible from public sources and proprietary sources, such as user-submitted surveys or website analytics. An algorithm that was trained with the calibrated dataset uses these data to predict the valuation as accurately as possible. This dataset is expansive and grows rapidly to include the most recent, relevant data for as many properties as possible.

Examples of property data used by an AVM:

• Property size (acreage)
• Home size (square footage)
• The number of rooms (bedrooms, bathrooms, etc.)
• General location (state, city, zip code, and sometimes even neighborhood)
• Home quality characteristics (air conditioning, pool, garage size, etc.)
• User-submitted information (ex. HomeLight’s simple home value quiz)
• Property price history
• Property tax valuation history
• Property historical sales information

How algorithms used in automated valuation models work

The algorithms that fuel an AVM use machine learning and regression techniques that take massive amounts of data in order to make accurate and reliable predictions. Regression and machine learning are buzzwords that you’ll hear often when talking about algorithms and mathematical models.

Regression is a mathematical technique that uses one set of data in order to predict another. When you have a calibration dataset, you can use both points of data (the data point you are trying to predict and the data points you are using to make the prediction) to create an equation or model. In this model, you can input only the second type of data, and it will output the prediction for the first data point that you are trying to predict.

Machine learning models go a step beyond regression by building a way for the algorithm to continually improve over time through a growing dataset. A machine learning algorithm can continually recognize patterns as new data comes in, and make changes to the algorithm to produce increasingly accurate and reliable predictions.

Why do different AVM tools produce different estimates for the same property?

You may have noticed that the same property can have different estimates depending on which AVM tool you use. Each of these tools takes both publicly-available and proprietary data and uses its own proprietary models to estimate values. Differences in the approaches to building the models, and the type of proprietary data, such as user-submitted data or website analytics, account for discrepancies in the value estimations.