Appraisal Statistics

[Joe Flacco]

Comparing smaller set of data with minimal differences is more reliable than using large sets of data with many differences.

 

[J Grant]

Pulled from....

If you selected an appraisal report at random, and asked the appraiser to provide actual support, using a recognized method or technique, for the GLA adjustment rate that he/she had applied, how many could produce that (without having to back into it now that the report was completed)?

Sorry but developing an adjustment from the comps themselves ( pairing them, extraction, line item sensitivity analysis ) is not PFA...and the appraisers using the simple, but time tested and accepted/recognized methodologies support them with CA, interviews, and experience...why is appraiser experience denigrated to mean nothing? If an appraiser developed an adjustment last week for a 2000 sf house with a bath in a 300k price range, and they are back in subdivision next week appraising a similar house, do they need to develop the adjustment from scratch all over again? Or can they rely on the fact that they developed it last week ?

The problem is with advent of computers, appraisers are supposed to be able to answer like they are a computer , or used one. IF you want appraisers to use computers for math results they can trot out, then declare it...but USPAP has that issue covered where they say if appraiser uses RA or stats the appraiser needs to understand them...which I doubt most could do without intense exposure/education in it...

Marginalizing appraisers and their human reasoning is the outcome of saying things like a few comps in SCA (that appraiser spent hours or days finding and vetting ) are nothing more than a weak, small set of stats and appraiser's reasoning counts for nothing or little..serves as the rationale to replace appraisals with computer products or a hybrid with an appraiser signature...

 

[DWiley]

You need to be schooled by Leased Fee about statistics, regression, and real estate data.

That would be interesting. Perhaps LF could teach me more than those crazy professors over in the Math Department at Vanderbilt University did

 

[Joe Flacco]

That would be interesting. Perhaps LF could teach me more than those crazy professors over in the Math Department at Vanderbilt University did

I bet you he could.

 

[DWiley]

Sorry but developing an adjustment from the comps themselves ( pairing them, extraction, line item sensitivity analysis ) is not PFA......

You totally missed my point. What percentage of practicing appraisers do you honestly think actually apply those recognized methods and techniques that you cite? And, in comparison, what percentage are simply using "the list."

If a big percentage were really using recognized methods, then how is it that a few years ago Fannie's data showed that so many were basically using $30/SF for almost all properties?

 

[Joe Flacco]

You regurgitate math people without real estate background and Leased Fee has real estate background and has studied statistical modeling in depth. As in he has done his own research.

 

[Joe Flacco]

Regression with coefficient of zero = paired sale. I think that is how he explained it.

It doesn't take a genius to understand that with large sets of data there is a lot of noise and is less reliable than isolating a single variable.

 

[DWiley]

I bet you he could.

No bet I started working in the appraisal business in 1982. I graduated from Vandy in '86 with a degree in applied math. Since I was working at an appraisal company, naturally my independent projects dealt with appraisal stuff, and I carried that through and applied it in my own practice for over 30 years. I think statistics and modeling has a nice place in our profession, but I also think that a lot of what is taught focuses too much on the math and too little on the application. I have, for example, examined a lot of regression courses that dealt with the math of regression exactly correctly, but then failed to illustrate its proper application to the average residential appraisal problem. That is why you see things like an appraiser using Stat Wing to build a model for a waterfront property, but having that model indicate that the site is not a significant variable (an example from an actual state board case), even though the cost approach showed that the land value was more than 50% of the total value. Some take those classes looking for a 7 hour or 15 hour solution to something that takes much longer to learn. They learn just enough to be dangerous.

 

[Joe Flacco]

No bet I started working in the appraisal business in 1982. I graduated form Vandy in '86 with a degree in applied math. Since I was working at an appraisal company, naturally my independent projects dealt with appraisal stuff, and I carried that through and applied it in my own practice for over 30 years. I think statistics and modeling has a nice place in our profession, but I also think that a lot of what is taught focuses too much on the math and too little on the application. I have, for example, examined a lot of regression courses that dealt with the math of regression exactly correctly, but then failed to illustrate its proper application to the average residential appraisal problem. That is why you see things like an appraiser using Stat Wing to build a model for a waterfront property, but having that model indicate that the site is not a significant variable (an example from an actual state board case), even though the cost approach showed that the land value was more than 50% of the total value. Some take those classes looking for a 7 hour or 15 hour solution to something that takes much longer to learn. They learn just enough to be dangerous.

Statistical modeling as a tool is good and fine but appraisers that don't even understand paired sales are not going to understand regression. If you are well versed in real estate data and statistics then you would understand that paired sales is as reliable as it gets.

 

[DWiley]

It doesn't take a genius to understand that with large sets of data there is a lot of noise and is less reliable than isolating a single variable.

Well, that cuts both ways in the real world. In the Stat Wing example I cited above, the resulting model could not/did not account for the impact of the site value because only similar waterfront properties were used in the data set. So, the impact of the site got spread out over the other variables (which are called that because they vary ) . A certain degree of variety is required in the data set to isolate any given feature, because without the variety it cannot be isolated. One cannot, for example, extract the contributory value of a two car garage using a data set composed entirely of homes with two car garages.