Regression for GLA Adjustment

[TerryRohrer]

If you are talking about the analysis of a single variable, there cannot be much difference in the meaningfulness of the results. If you pair 4 or 6 or 10 sales in all possible combinations, or plot the data and show the equation of the best fit line (simple regression) there can be no appreciable difference in the results. The power of regression begins to emerge when other variables are added, and as more data is added. There is far more insight available through this tool than most appraisers realize, but most won't do the hard work of getting even a rudimentary grasp of the process. In one conversation, another appraiser assured me he had given it the old college try, even took a class. But he got hung up on the explanation over how to export data from his MLS. I guesstimate that is as deep as many appraisers have delved into the methodology.

 

[alebrewer]

Can you create meaningful regression results using only 4-6 properties? Or using 10 properties?

My reports typically have 4-6 comps ( these days 5-6) but in addition I analyzed and review more sales and listings in addition, how many depend on the assignment. So although the 4-6 comps are what show up on the grid, 8-16 additional sales /listings/ pending contributed to the analysis . (typical number, could be more or less after elimination of others )

The typical buyer does nto run regression, so staying closer to their thought process and analyiss more closely can replicate/understand their motivations in the market -USPAP also has a section (paraphraisng it here ) where the appraiser , if they use statistics or AVM or other computer program methods should be able to replicate results if asked - how many can do that ?

If a person just feeds the data in and lets the machine do the work, where is the understanding -(not saying that represents any one person here, more of a general question )

Yes - in fact, let's say you analyze 6 sales in your SCA. If you've correctly identified the elements of comparison, then cite those elements of comparison as the independent variables in a multi-variate regression (with Sales Price being the dependent variable), your 'regressed' adjustments should be nearly identical to the adjustments you applied via sensitivity analysis, grouped data analysis, etc.

 

[CGinMN]

Start by reading The Valuation Analyst by David Braun. He explains it in the easiest of terms. That was my entry into MLR. I don't use it in every assignment, but I am a better appraiser for understanding it, and generally how valuation modeling works.

 

[TerryRohrer]

What do you do when all of your statistics come up with a result that you know is incorrect based on your expertise in the local market? experience is everything.

In my experience, that is almost never the case. Once you set up a model and view the results, you can almost always think through the results in the context of the dataset used, and understand why the outcomes are as they are

For example, I often see dramatic, definitive assertions that an age adjustment is never warranted and differences in age has to be adjusted elsewhere, commonly in condition. Yet, I almost never find that differences in age are not a significant variable.

That reaction is highest (in gross dollars) the newer the properties in the data set, and wanes to almost nothing in older properties. Why? Well, I see, and see others reporting the same, that new construction sells at a significant premium over existing. The disappearance of that premium is obvious in regression results. But at 50 years, and moreso at 80-100 years, it all but disappears. Why? Because you can have houses of the same age side by side selling for vast differences in price due to condition and updating.

There are only a handful of variables that are almost always remaining in my final models, and those vary only slightly (in number of variables) across a wide swath of property characteristics. But once you have seen them enough, you know the results are sound. That leads to huge potential for exploration of other variables' impacts on value. Busy street, pool, view, outbuildings, most take minutes to add to a model and rerun.

Of course, I was accused yesterday of lacking any understanding regarding the markets for cookie cutter tract homes, so what do I know? Well, there is not much of that here, but I can assure you, if you always have 50 comps identical to the subject, 1) regression won't work (it can't run if there is no variation), and 2) no lenders will be hiring an appraiser for those soon. A simple average will be suitable.

 

[chad hampton]

Maybe, I just know enough about statistics to know that I can get a sample set to say pretty much anything I need it to say. This was always a profession where your local experience counted for something. I don’t think it’s wise to get away from that. real estate is and always will be local. I realize that doesn’t fit the narrative to funnel money into the big corporations.

 

[TerryRohrer]

Maybe, I just know enough about statistics to know that I can get a sample set to say pretty much anything I need it to say.

If you don't think that can be done with any methodology you employ, you are missing a great deal. Do you actually believe those who intend to mislead, whatever their weapon of choice, can't do that with paired sales? Or comp selection? Or false statements? My comments are offered with the underlying assumption that the reader is an honest appraiser. Nothing I offer should be read or relied on by any others.

But while you refuse to understand how to rely on "statistics" to improve your product, others have concluded that their results from same are sufficiently reliable enough to replace you in a very high percentage of cases.

 

[RCA]

My issue with statistics is it needs a volume of properties to work, and a number of those properties will not be comps for the subject and then the question is did that skew results....

Using the comps a buyer really would consider is about quality vs quantity. Plus buyers do not make a chart or graph, they have a rough idea of what they want to spend- it is not that hard to study the patterns and figure it out from the activity itself, verify with RE agents and other support.

You will (almost) always need about 5*(the number of relevant features) property sales to get a sufficiently reliable price model. However, those transactions can go back more than one year - if your regression software can properly model price changes over time, with up and downs. So, you will need a non-linear regression tool like MARS to do that. You can also go beyond the neighborhood - if your regression tool can adjust for location - and so it will have to be able to handle non-linear relationships and adjust for area along with everything else - MARS can do that.
You might ask what good it does to go back more than a year when your comps going into the sales grid can only be in the past year or in the subject neighborhood. Well, by going further out in all directions, you get a larger sample of homes with various kinds and combinations of features that allows MARS to obtain increasingly more detaled levels of intelligence. In the end you get a smart model that can model nearly all important differences - and apply that to a subset of transactions that are required by lender/GSE guidelines.

However, by restricting yourself to 3-6 transactions, you are automatically biasing your valuation to the random perturbations showing up in that small sample. Your value opinion is then subject to the bias imposed by the small subset you have chosen - it is likely not reliable. -- This should be old information, well established many years ago. It is basic (advanced) statistics.

 

[Fernando]

You will (almost) always need about 5*(the number of relevant features) property sales to get a sufficiently reliable price model. However, those transactions can go back more than one year - if your regression software can properly model price changes over time, with up and downs. So, you will need a non-linear regression tool like MARS to do that. You can also go beyond the neighborhood - if your regression tool can adjust for location - and so it will have to be able to handle non-linear relationships and adjust for area along with everything else - MARS can do that.
You might ask what good it does to go back more than a year when your comps going into the sales grid can only be in the past year or in the subject neighborhood. Well, by going further out in all directions, you get a larger sample of homes with various kinds and combinations of features that allows MARS to obtain increasingly more detaled levels of intelligence. In the end you get a smart model that can model nearly all important differences - and apply that to a subset of transactions that are required by lender/GSE guidelines.

However, by restricting yourself to 3-6 transactions, you are automatically biasing your valuation to the random perturbations showing up in that small sample. Your value opinion is then subject to the bias imposed by the small subset you have chosen - it is likely not reliable. -- This should be old information, well established many years ago. It is basic (advanced) statistics.

Of course the larger number of transactions, the greater the reliability. I normally try to get at least 20 transactions before I feel comfortable.

 

[RCA]

Maybe, I just know enough about statistics to know that I can get a sample set to say pretty much anything I need it to say. This was always a profession where your local experience counted for something. I don’t think it’s wise to get away from that. real estate is and always will be local. I realize that doesn’t fit the narrative to funnel money into the big corporations.

Well you are off base here. You shouldn't be creating statistical models - if your goal is making them say whatever you want. OF COURSE, you can bend statistical results by select subsets of data that will lead to the desired outcome. You can lie with statistics.

For good work, you have to be highly objective in selecting the data, setting the parameters, and ensuring that the results have sound logic.

There are, nonetheless, certain areas where you will find no sense in statistical results because of factors like the following:

1. Terrible data quality. The sales agents entering their data over the base years are either sloppy or simply do not have the means to get the data other than through hearsay and guessing.

2. Sporadic and infrequent market activity with informal cash-only sales with many unrecorded concessions, trading and the like. Some beach communities up the northern California coast have been like this at times in the past.

 

[RCA]

Start by reading The Valuation Analyst by David Braun. He explains it in the easiest of terms. That was my entry into MLR. I don't use it in every assignment, but I am a better appraiser for understanding it, and generally how valuation modeling works.

Not bad as appraisal statistics books go. And it is at a level some appraisers can understand.

If I wrote a book, only a small handful of appraisers could get anywhere with it and I am sure it would be very slow going.

I am doing that --- but I am in no way writing it for current appraisal community - but the future. It will be complex, immense and cover many areas. And it will be done on my new website.

Right now at this moment I am working on:

"A System To Automatically Generate Human Understandable Descriptions of Complex MARS Models"

This could take a MARS model like the below (it is contrived for the purpose of testing) and simplify it into a number of sentences, using an ANTLR 4 Parser, based on Java code - which runs on Linux or Windows (and it handles both 1 and 2 way/degree models) [ and think about how much a person would have to know, the skillset required, to deal with this kind of problem? (1) Extensive experience with appraisal and studying MARS models for many different kinds of areas, (2) programming (in Java on Windows and Linux, maybe with Docker), (3) parsers (advanced Antlr 4 with listeners and visitors), (5) mathematics (not particularly advanced but nonetheless complex with many moving parts) and (6) Advanced Statistics (MARS):

(Intercept) 1994618.995226
h(21-Age) 18134.284216
Age * h(Baths-3) -13742.01604
h(29-Age) * Baths -15319.87281
h(Age-29) * Baths 2207.79385
h(Age-29) * Beds -2363.72638
h(Age-29) * FrPlcNbr -1707.98934
Age * h(GLA-2700) 3.34516
Age * h(GLA-2175) 54.76422
Age * h(2700-GLA) 40.45922
h(29-Age) * GLA 70.44614
Baths * h(DaysOffMkt-2446) -1856.25314
Baths * h(DaysOffMkt-1542) 720.90125
Baths * h(2446-DaysOffMkt) 121.76920
Baths * h(GLA-2700) -667.54114
Baths * h(GLA-2175) -246.22563
h(DaysOffMkt-339) -759.797818
h(DaysOffMkt-975) 843.678601
h(DaysOffMkt-1560) -403.716656
DaysOffMkt * h(1-PoolStatus) -169.77454
h(DaysOffMkt-1542) * PoolStatus 0.58995
h(DaysOffMkt-2446) * PoolStatus -182.74645
FrPlcNbr * h(2700-GLA) 603.22737
h(2-Garage) -86296.145126
h(GLA-3040) 317.051132
h(3540-GLA) -197.630858
h(GLA-3924) 1296.744498
GolfCourse 124657.567723
OceanView 198593.112427
h(Latitude-36.955) 93569930.354
h(36.968-Latitude) 72358503.742
h(Latitude-36.968) -128469121.497
h(Latitude-36.952) * h(Longitude- -121.882) 9445111406.771
h(36.968-Latitude) * h(Longitude- -121.875) 3437590452.22
h(36.968-Latitude) * h(Longitude- -121.881) -3712978768.509
h(Latitude-38.64) -6877270.02085
h(Longitude- -121.882) 9445111406.771
h(Longitude- -122.88) -34011087.24943
h(61420-LotSize) -18.76882
h(LotSize-61420) 0.22612
h(1-Pool) 782261.64137
h(Pool-1) 19970.69823
h(2-Story) -68741.56691
h(Story-2) -228532.34087