Automated Valuation Modeling (AVM) – Single Family Homes
Automated Valuation Modeling (AVM) is a term given to home price estimates that utilize statistical or mathematical applications to assess the value of real estate. Single-family residences and condominiums are the subjects of these high-speed estimates that take place in less than a minute. At present, there are some fifteen models, and perhaps a few more, that exist in the United States. Most of these applications have been in the market for a minimum of seven years and some, for as many as twenty years. There are also some new wannabes websites who have brought into question the validity of AVM with their wild numbers. AVM, properly applied, is a very viable tool in today’s real estate market.
AVM’s have been used by the mortgage lending industry as a means to understand values so that lenders can 1) save some of the costs related to appraisal and 2) manage the risk of all of the loans that they originate or sell into the investor market. In other words they want to manage their risk and use automated information tools to help them streamline their processes. AVM’s capture sales data in the area surrounding a subject property, analyze it, throw away “outliers” that don’t conform to the subject property and then the model calculates a home price estimate.
Most models have been tested and re-tested through the years. In fact, most large-scale lenders conduct what is known in the industry as a “data bake-off” once or twice each year to see which models are best. Lenders want to know what companies are going to the effort to improve and refine their models. From there, they re-define which models will be used for that upcoming time period to help them automate their processes, and improve accuracy. It is a costly, time-consuming test and only certain companies even choose to compete in these tests. Those that don’t compete simply aren’t given the opportunity to earn that lender’s business, so most choose to compete. HomeSmart Reports uses AVM results from only those companies that compete in industry testing.
In conducting their tests, lenders most often use a sample of their own “closed loan” portfolio and also a sampling of closed sale transactions from publicly recorded data. What better way to validate the accuracy and the “hit rate” of the models? Accuracy relates to how far the AVM estimate is from the actual transaction amount. It would extremely rare that a model would predict the exact price of the prior sale since it analyzes so much information to make its estimate. However, the AVM’s consistently closest to actual sales prices are deemed to be the industry leaders. The variance in the actual sale price or appraised price and the AVM result is measured on each property. Hit-rate is a measure of how many times the model is able to come up with a value estimate.
Geographic coverage of data is a discussion all its own. Major data companies (and there are only a few of them) go through a very painstaking process to ensure that they acquire and standardize as much data as possible across the country. It is a very expensive process. These companies, by and large, supply the data that drives AVM engines. Without good data, AVM’s are less accurate. In the computer world, there is an old adage, “garbage in, garbage out”, or GIGO. AVM companies understand this and so they use the high quality information acquired by these data compilers to “fuel” their statistical engines. Most of these sources are from public records though some AVM’s claim to have relationships with local and regional (not national) Multiple Listing Service (MLS) providers. This means that you can get varying quality from one region to the next in terms of accuracy.
Data quality depends on content, coverage and currency, the so-called three C’s in the information compilation world. Content refers to available data, such as property characteristics like bedrooms, baths, square footage, lot size and a host of other property descriptors. Coverage refers to the geographic scope of the data. As an example, if you have the “best model” provided you have right data to fuel it, and that data only exists in ten counties in the country, then your model is only the best in those ten counties. If your model can’t adjust, or hasn’t been adjusted to accommodate the varying availability of data elements, then your model is scalable to national levels. If it isn’t scalable, then it can’t accommodate the needs of a national service.
Many models have been built to accommodate a minimum number of data elements to calculate a statistically accurate value. If there are more data elements available, they are used by the model. If not, the models revert to the minimum required numbers. There are other types of models and those will be briefly discussed. The final C is for currency, which partially dictates how up-to-date the sales information is for the model estimation. For example, if sales data is 6 months old, the model is likely to be less accurate than a model using three-month old or one-month old data.
Most of the higher quality AVM’s also factor in market trends into their application. That is, is the market moving up, down, or staying static? Applying a trend-factor or growth/decline index helps smooth-out anomalies in the value calculation (or market volatility) and creates more accurate values. AVM’s generally do not have the most recent sale in the neighborhood. Due to the compilation process, which often requires weeks to capture data, standardize and apply it to the proper systems, sales data may appear to “lag”. Homebuyers and sellers have been conditioned by the real estate industry to expect the most recent sale from around the corner to totally dial-in on the value. This is not a bad thing, but it sometimes causes them to question the validity of an AVM if the most recent sale is not among those listed. One sale more, or one sale less, will not have a statistically significant influence on the calculation that estimates value. An AVM will get you very close to a value, but don’t expect them to give you an exact value.
It is important to step back for a moment and understand that with AVM technology, models cannot “see” the particulars of a property such as a well-manicured yard and a home with fresh paint, or a property that is run down. AVM is not a silver bullet to home valuation. An AVM is designed to deliver a specific estimate of value, with a high and low range and provide a confidence score relating to is accuracy. The confidence score will depend on the number of data points (in this case home sales) that were available for the calculation. Generally speaking, the more data points available to calculate the value, the higher the confidence score. The fewer the number of sales data points, the lower the confidence score. Buyers and sellers should use AVM as a “benchmark” to compare the value of their property against other sales in the area and against homes that are currently available on the market. Either that, or totally trust your real estate broker if that’s what you choose to do.
Another type of AVM is known as a Repeat Sales Model, or Paired Sales Model. A repeat sale occurs when the same property sells once and then sells again. When the difference in values is measured against the time between sales, a growth index for the area can be developed. As with other models, there needs to be a sufficient number of data points to develop the index. Once the growth index for the area is established (and the index changes over time), the index can be applied to a prior sale amount of a property to estimate its current market value. This does not take into account value differentials that may occur due to additions, remodels or changes to a property.