Back in 2000 the American Appraisal Institute stated that ‘the market needs a more rational, objective, scientific basis on which to make decisions, performed by professional individuals competent both in appraisal theory and practice and in computerized data analysis and reporting’.

They envisaged a new approach, based on a blend of appraisal theory and statistical analysis. By combining user interaction leveraging photos with AVM technology, outstanding results can be delivered and used for applicable loan types or simply applied as a quality control tool.

The subjective measures required to optimise AVM results fall into either nominal or ordinal categories. Examples include views (yes or no) or street appeal (1 to 10 rating). To capture and use qualitative data in automated models is possible, but it does create a wide range of challenges.

Quantitative data is more objective, is generally easier to measure and more precise than qualitative data. In most cases, it is easier to capture larger samples or even population level data when looking at land size or typical attribute data like bedroom count.

Applied to AVMs, we can employ various models that just use quantitative data such as land size and bedroom count. In many cases, the results will be adequate.

Qualitative data when available can prove valuable and have a significant impact on the valuation modelling process. We often find this to be the case in locations with a wide range of house values, which may be attributed to factors other than land or house size.

However qualitative data needs to be correctly ranked or systematically treated in order for it to be reliable. The measurement system used to collect qualitative data can be validated. For those of you that have learned a little about Six Sigma would be familiar with the DMAIC and DMADV process, which often uses a test known as Gage R & R.

Applied to our example, a ‘street appeal’ ranking of 1 to 10 would need to be collected using the same standard for all houses used in the AVM sample. If the same person was employed to capture all of the measurements, we would potentially perform well in the Gage R & R test.

But if we employed 2 or 3 different people to capture this data, each with different tastes and standards, we would likely find a level of variability in the data that would be attributed to how we collected the data, rather than variation applicable to just the houses themselves.
Valuers have long used a simple method that resolves this problem.

By carefully selecting 3 or more matching properties in terms of size, sale date and location, a simple ‘comparison to subject’ measure can be applied to account for any qualitative differences.

By doing things this way, the data sample can be adjusted at the time of creating the valuation rather than relying on an entire database of various quality ratings. For example, the prices of the 3 properties may all fall within 10% to 15% of one another. Most of this residual difference in price may then be attributed to the qualitative differences between the houses. We may expect the most expensive house to be ‘superior’ and the least expensive to be ‘inferior’ to the subject.

This ‘user assisted’ approach can be split into 2 distinct steps in our example. In step 1 the computer would use a set of algorithms primarily leveraging quantitative data. This would include sale date, location, street type, dwelling type, land, bedroom, bathroom and parking data to find the best matching comparable sales.

Emulating how a valuer would perform this step, it can return 5 or 6 comparable sales within a few seconds in an entirely ‘rational, objective and scientific basis’. In a true automated model (AVM), the price estimate could be calculated at this point, treating all 6 sales as equal or ‘comparable to subject’.

Step 2 would then be considered the ‘assisted model’. For example, a user may remove unwanted comparable sales and have the computer automatically replace it with the next best matching property. Once the final comparables are selected, a user can leverage photos to adjust for the qualitative differences between each of the properties in comparison to the subject. Rather than relying on qualitative data, the user achieves the same objective using photos and adjusting ‘comparison to subject’ much in the same way as a valuer does it today.

The result will ultimately be just like the Appraisal Institute forecast, delivering ‘the most efficient econometric blend of economic (appraisal) theory and statistical analysis, on the foundation of mathematical relationships’.

Products such as Google Street View have helped make this possible via the web. The model described above is now available in Australia and is attached as standard with our pdslive subscription product (unlimited use and no transactional fee).

If you like to experience it for yourself, a free 7-day trial of our beta version of PriceFinder v2 is now available. Just email kent@pdslive.com.au

Story by Kent Lardner for Lending Central