Appraisers' Free Forum

http://www.fhfa.gov/webfiles/15151/10-30-09%20FHFA%20Default%20Risk%20Evaluation%20Report.pdf

I'll comment later.

The essential tradeoff between AVM estimates and traditional appraisals is well
known. It is hard to pervert AVM methods merely to “confirm” the value estimates of
market participants, but it is also hard to value aesthetic attributes of properties using
AVM models.

No it isn't. All you have to do is skew the data by eliminating those "undesirable" data points. But, yes, it is not easy to value aesthetic attributes with an AVM... in fact, it might be impossible. Still, we have appraisers out there saying that no one should ever use any subjectivity in appraising.

Naah. Rank the aesthetic data point according to a predefined system. Define the system before you go, and revise it as necessary. Using "view" as an example, your rank scale might be from 1 to 5, with 5 being the best. The very best view would be a 270° unobstructed view of a lake, from a promontory point. The very worst view (a "one") would be large trees and your neighbor's outhouse. An average, or 3, would be a 90° view of the lake without obstructions; a 2 might be a peek-a-boo view around trees or other houses; and a 4 might be a 180° unobstructed view. Now, using that predefined scale, rank your comparables' views as well as your subject's. In your regression model, introduce a discrete variable carrying that rank. The coefficient of that element will determine the emphasis it gets. That is, if the coefficient is $9,000 per rank point, the value would be determined by rank x coefficient. In the example, if you subject had a 4-ranked view, then view would account for $36,000 of its selling price.

The problem with this is buyers do not always react to such a feature spectrum in a continuous manner. More "quantum" than "continuum". In recent days, with the market downturn, these quantum jumps in appeal for various aesthetic features (negative and positive) have magnified. Your model would have to be continuously tweaked to provide any meaning conclusions. The constant tweaking would nulify the cost savings. As useful as the AVM might be for analyzing hard data (GLA, etc.) its just not going to work well for aesthetics.

Jim Plante wrote:Naah. Rank the aesthetic data point according to a predefined system. Define the system before you go, and revise it as necessary. Using "view" as an example, your rank scale might be from 1 to 5, with 5 being the best. The very best view would be a 270° unobstructed view of a lake, from a promontory point. The very worst view (a "one") would be large trees and your neighbor's outhouse. An average, or 3, would be a 90° view of the lake without obstructions; a 2 might be a peek-a-boo view around trees or other houses; and a 4 might be a 180° unobstructed view. Now, using that predefined scale, rank your comparables' views as well as your subject's. In your regression model, introduce a discrete variable carrying that rank. The coefficient of that element will determine the emphasis it gets. That is, if the coefficient is $9,000 per rank point, the value would be determined by rank x coefficient. In the example, if you subject had a 4-ranked view, then view would account for $36,000 of its selling price.

Yes, that would work (as well as the technology works). However, the data would have to be available. I just saw a beautiful, 6,000 Sqft.+ house overlooking a bay at Tampa. Full, beautiful, 270 unobstructed view, right on the beach (and it's for sale at a lot less than last year, I'll bet). If that was your data, the view would be very best. That's out the back side of the house. Front side faces a sub-division street, but there is another small bay just beyond that with a large power plant in full view. Your regression model would miss badly based on your paragraph... you'd have to know and account for the front to properly value the house.

And, Skibs is right. Some buyers want a large lawn. Others don't want to mow that much. Regression will miss every time unless it accounts for the subsets. And, I haven't seen an AVM yet that does, although I know it is possible to do so.

Steve and skibs,
Your comments draw a bright line underscoring the fact that statistics is merely a tool, and not an oracle. My comment was geared not to an AVM that would eat anything and value it, but instead I had in mind modifying a model for a particular appraisal. And that's what one has to do when using these tools: Fine-tune them for each report. The model may be automated, but my automated transmission still needs me to steer the car.

What do you do about the lake that is in the process or just did dry up?

How about the lake is eroding the land and the water is closer and closer to the house.

Tel, that is why I said it is not easy to value aesthetics with an AVM (and might be impossible). There could be mathematical value points assigned to the kind of events you are describing... if there was enough data. But, sooner or later, you are going to run into something that simply cannot be valued that way. At that point, you have to use your head (and experience) to decide with the aesthetic factor is worth.