Avm Risk? Avm Accuracy Above That Of Appraisals

[George Hatch]

"Sale Price" is like a grade the market gives a property for quality, condition, functional utility, view and other "intangible" features, after you take away the value for hard features like GLA, lot size, garage size and so on.

I believe, with a good model you can have a very accurate valuation - but the market itself is not that "efficient" or accurate. Sellers, buyers and agents are typically not that knowledgeable or judicious. Also, there can be pressures to sell fast or not sell at all. A seller can say: "Yes, I would be willing to sell if I can get at least $X" and be willing to waste everyone's time trying to sell it for that price; only to cancel in the end. Or the Seller can say: "I have to sell this house in 4 weeks for whatever price I can get." Or the agent handling probate, can do a poor job of showing and hand the house off to a friend of a friend for 15% less than what he could have sold if for. In these latter cases, the owner (or other parties as the case may be) would be judicious to get a good appraisal ... but that requires public confidence in the appraisal process, which is lacking.

The definition of Market Value speaks explicitly to the actions of the typical buyer and seller. Not the smartest buyers and sellers.

There are other definitions of value out there that we could use and there's nothing stopping an appraiser and intended users from devising their own definition of value upon which to base an appraisal (or valuation). The big point is getting the appraiser and their intended users on the same page as to what that valuation does and doesn't mean.

If/when the AVMs and data science folk are unhappy with "typical" and want to use a more rigorous definition that better suits what they consider to be the strengths of their applications then maybe they should find or devise and promulgate the use of that different definition of value.

Perhaps something along the lines of "Data Science Value is the price at which a buyer would pay and a seller would sell if they were as smart as the valuation model".

I'm mostly not kidding.

 

[J Grant]

Around 1994/1995 BofA was testing an AVM (don't know if it was called that back then) system for equity loans....

The appraisal department had staff appraisers and AVM appraise the same property....
The AVM didn't follow all the guidelines that appraisers followed (example we stayed with city borders and sometimes the AVM went outside)....

But at the end of the day many AVM values were similar to human provided values....
Enough so that I could see the writing on the wall....
Less volume and lower fees....
But it seems it took a couple of decades to gain traction....

That's cool, but if the AVM's were so good, then how come regulators several years ago change the regs so that an AVM alone can NOT be used to UW a loan ? Such as Heloc, equity, the loans they were used for. It changed to at minimum an evaluation with an ext property inspection. It seems to be evolving toward rather than an evaluation, the hybrid desktop appraisal is the choice with firms gearing up for that.

No matter how much data there is, an AVM has a limitation of how "good" it can be. The reason why fannie;s own y value model/ AVM's are better, ( thus used for appraisal waivers,) is, ironically, because of fannie;s vast appraisal database.. A property that qualifies for a waiver had an appraisal done within X past years .

 

[ucbruin]

That's cool, but if the AVM's were so good, then how come regulators several years ago change the regs so that an AVM alone can NOT be used to UW a loan ? Such as Heloc, equity, the loans they were used for. It changed to at minimum an evaluation with an ext property inspection. It seems to be evolving toward rather than an evaluation, the hybrid desktop appraisal is the choice with firms gearing up for that.

No matter how much data there is, an AVM has a limitation of how "good" it can be. The reason why fannie;s own y value model/ AVM's are better, ( thus used for appraisal waivers,) is, ironically, because of fannie;s vast appraisal database.. A property that qualifies for a waiver had an appraisal done within X past years .

Human genetic modification may be doable but not yet acceptable....

Regarding your question, I haven't a clue....

 

[RCA]

The definition of Market Value speaks explicitly to the actions of the typical buyer and seller. Not the smartest buyers and sellers.

There are other definitions of value out there that we could use and there's nothing stopping an appraiser and client from devising their own definition of value upon which to base an appraisal (or valuation). The big point is getting the appraiser and their intended users on the same page as to what that valuation does and doesn't mean.

If/when the AVMs and data science folk are unhappy with "typical" and want to use a more rigorous definition that better suits what they consider to be the strengths of their applications then maybe they should find or devise that different definition of value.

Perhaps something along the lines of "Data Science Value is the price a buyer would pay and a seller would sell if they were as smart as the valuation model".

Look at yourself as a buyer and as a seller, put yourself in that situation and imagine how you would act. Many of us have bought homes. In many cases, we are quite knowledgeable and smart. Smart enough. In many cases, the seller and/or buyer are playing hardball. In the end, the buyer has to decide whether to pay $X for the property, where the seller simply won't go any lower. It may be more than he would like to pay for the property. But nothing else around to buy in the desired location comes close to matching the requirements. I would say, generally, in 80% of cases (Pareto's Law), that sale price is a fairly good indicator. But, there are certain kinds of sales, by their very nature, you really want to exclude. In my experience, probates are often gateways to cheating by agents to unload a home for below market value. You have to look at these sales closely. Shorts are typically under pressure to sell to avoid foreclosure. REOs are kind of their own market and need to be looked at closely.

The valuation model just mimics the market. It searches for a conditional average. If I have a linear model that says GLA sells for $400/sf for homes between 1000 and 3000 sf plus a base price based on other features and happen to have two comps with the same base price at 1500 and 1800sf and my subject, which just happens to come out with the same base price but has a GLA of 1650 sf, then my value will be exactly between the value of the two comps. Someone can figure this out and inform the seller and buyer, and that may well be the agreed upon price. But if another buyer comes along, he may be willing to pay a bit more, even if he is informed of the model, because there is simply nothing else around that quite fits the bill for that price. The model can never by that perfect. For one thing, it is always based on past performance and trends, and the world is constantly changing, trends are continually changing. For another thing, it deals with averages and "expected" values, never proclaiming to be able to exactly predict future outcomes.

So, you can't say that a smart seller and buyer will abide by a particular model. On the other hand, if the seller (owner) has a model from a competent appraiser that says his property is worth $1.3M and his agent hands him an offer to buy the house for $900K, then he would be smart to find out what the hell is going on with the agent.

 

[Joe Flacco]

If you have homes between 1000 SF and 3000 SF, and all else is equal, then your $/SF is always going to be a lot higher for the 1000 SF home than the 3000 SF home. Mainly because of the value of the land is a large portion of the $/SF which is the same regardless of the size of the house.

The only thing $ / SF is good for is comparing houses that are the same SF. Then it tells you one property is $x / SF better or inferior to others for whatever reason.

 

[RCA]

I think the amount of "intangible" features has decreased significantly over the last 10-20 years. A lot of sales which would have been considered outliers in the 90's are no longer outliers. These former intangible features are not always databased but available for consideration.

Interesting, I think the opposite. The "intangibles" or to be more exact, contributors to value that are not numerically and accurately measured in the MLS databases, seem to be growing, as the world becomes more complex. I can give you significant contributors to value in my area, that the GSEs have long failed to realize, such as the ability to rent out rooms in a home. You can get $3K/month by renting out three bedrooms, a bath and sharing the kitchen in this area. Invariably, you want a home that minimizes the interaction between the boarders and the owners. Homes are very different in this respect. The ability of seniors to retire in their homes, depends heavily on this factor. Then, of course, there is technology. There are perceived/real threats that vary by location and change with time (increased earthquake, fire, crime hazards). Generally, as well, as neighborhoods get older, they become more complex through gentrification, upgrades and other factors.

 

[Joe Flacco]

Interesting, I think the opposite. The "intangibles" or to be more exact, contributors to value that are not numerically and accurately measured in the MLS databases, seem to be growing, as the world becomes more complex. I can give you significant contributors to value in my area, that the GSEs have long failed to realize, such as the ability to rent out rooms in a home. You can get $3K/month by renting out three bedrooms, a bath and sharing the kitchen in this area. Invariably, you want a home that minimizes the interaction between the boarders and the owners. Homes are very different in this respect. The ability of seniors to retire in their homes, depends heavily on this factor. Then, of course, there is technology. There are perceived/real threats that vary by location and change with time (increased earthquake, fire, crime hazards). Generally, as well, as neighborhoods get older, they become more complex through gentrification, upgrades and other factors.

I think renting out rooms in most high density places with zoning is illegal use. If it is a legal use contributing to value then that is a zoning issues that contributes value. You don't have two houses next door where the value of one property is more than the other because you can rent out rooms.

Just because a property has a difference that is not numerical or databased does not mean that it is intangible. There is so much data available now with a few clicks of the mouse. How much data is in photos that is not databased? That's not including all the websites like interactive zoning maps and permit applications sites.

I agree 100% with what you say about as neighborhoods get older, they become more complex. That is basically the main issue that the ibuyers are having with the northeast and other high density locations and why they say it is unpredictable and singular.

 

[Joe Flacco]

Not only do we have current photos for many properties, probably have sets of photos from 7 years ago and 15 years ago that can be compared with the current photos to see how the property has changed over time. There is a lot that we know about properties now that is not numerical or databased that answers a lot of what was unknown or intangible in the past. Back in the day in the 90's I don't think appraisers even knew if a house has hardwood floors or carpet.

 

[George Hatch]

Specifying and calibrating a model to accommodate a narrower range of variables will be easier and more effective than trying to build a model to accommodate a wider range of variables. The problem is driving enough assignments into a given model to economically justify its development and maintenance.

 

[Joe Flacco]

No it's not. Recognizing that there are many more variables than what is databased and finding the handful of comps that are the most similar is the most effective. Crunching large sets of data for narrow range of variables will never be more effective than finding the handful of most similar comps.