TAF and USPAP - great analysis

[RCA]

Over the past 15 years I have seen many examples of situations like that. It has been my experience that the root cause is not the underlying methodology. Rather it is most often due to the failure of one, or both, to execute the methodology cited or claimed.

I strongly believe that both Fannie Mae and Freddie Mac are composed of people at all levels who are absorbed in doing little more than keeping their jobs and making a good salary for as long as they can. They are lucky to work for such solid corporations and must be happy with themselves.

- Which is probably the best explanation for why they have their head in the sand on an almost permanent basis, see no further than necessary, learn no more than necessary, do no more than required, and basically most likely focus on only what they need to do to get along with their management hierarchy and their peers where they work.

How can anyone object to that?

Social Responsibility? --> That's a good question I have never looked into. I don't have time. But I would guess they are largely shielded by Federal Government mandates and all kinds of built-in rules and regulations.

As a result, you are wasting your time talking to them. If you could find someone at these corporations to seriously listen to your recommendations, they would not be able to do anything, even if they were high-level managers - without risking their jobs. - And that is not going to happen.

 

[Non Sequitur]

Jeff Bradfod, at today's Zoom AI Conference in Monterey, CA, said he is coming out with a new product called "Nighthawk" that will also sell regression models for neighborhoods. I recommend using R/earth (or MARS), - he isn't being that specific, although anything else would be a black box. He says he has Ph.D.s building the models, which is like him. I tried to sell him on integrating MARS into his software in 2007-8, but he passed. OF COURSE, getting it off the ground would have taken much effort and many years. He has been working on and off of AI photo analysis since about 2015.

Well, he has always had a not good reputation for carrying projects through to the end. Maybe now he will, as he has always had some cooperation from Fannie Mae. He won't answer any specific questions about what he is doing, but I guess he has money somewhere to hire a few programmers and statisticians. I think maybe I know who.

This is the scenario best I can make out:

1. NightHawk will generate reports in the new Fannie Mae format, expected to arrive sometime in 2026. Maybe later.

2. NightHawk will provide models by neighborhood. We don't know how often they will be updated or specifics. The models will be created by Ph.D. statisticians who are not cheap. Maybe he has the political weight to change the laws so statisticians can do appraisals without an appraisal license. -- Just one problem of many to be sure.

3. There is a liability problem. Other companies like Zillow and House Canary could also essentially sell such models -for thousands and thousands of appraisals across the US. They have thought about it, but have not carried forward for sundry reasons. Some such appraisal models would be messed up, absolutely for sure 100%. -- And then the question is, who will pay the price for the ****-ups. Jeff wants the appraiser to take full liability. He says he will have courses to teatch the apppraisers what they need to know. Good luck with that.

4. A good MARS regression model will predict all the measurable variables, such as GLA, room count, lot size, etc. In the SF Bay Area, that is 70-85% of the value -of the subject. The appraiser gets to determine the residual for the subject, which is added to the MARS estimate. So, if there is a significant problem with value, -> that points to Bradford Software, you can be sure they will be sued, if anyone. If only the appraiser gets sued, it won't take long for appraisers to become very shy of the software. In any case, there will be enough cases where the models will not be good enough for non-conforming homes- and then what to do? Jeff hasn't figured that out - at least, he didn't have an answer today. Of course, if appraisers sign an agreement to take full responsibility for the accuracy of the models, Jeff is off the hook. Oh yea. - Unless they start highering math and STEM grads to do the work, I think he will have problems.

Anyway, there isn't any value-added service appraisers can provide in this new system unless they are MARS/Appraisal savvy. Given that Stephen Milborrow has been working for 20+ years on R/earth, I don't think Jeff's programmers have much chance to create their own. Stephen Milborrows program is open source. If Jeff uses the code, he must provide it for free to the public. He could get a license from Minitab for their MARS - which doesn't work too well with their scripting languages, not as good as R in any case, but still good. Only one individual license is $16,000/year. I am unsure what small fortune he will pay yearly for several thousand appraisers.

It would be a good thing if Jeff could get his NightHawk to float. It will make headway for all who want to use advanced statistics. -- But we see what happened to Redstone and other projects he has started. The Glassdoor comments give some insight (they aren't from me, btw). - Jeff would instead go on a vacation to Russia or some such place rather than finish his startup projects if past patterns of behavior mean anything. But, on the other hand, maybe he felt the time wasn't ripe for change in the past, and now it is. Hmmmm.

A representative from Fannie Mae was present at the conference to discuss their great new report system. I let him know what I thought about their great progress in the past but that they still have their head stuck in the sand. -- No, I didn't say that, but kind of, the moderator of the Zoom conference was trying very desperately to cut me off. But I said enough. And somebody said: Well, he won't get any more Fannie Mae appraisals! Ho hum.

Basically, he (and many others) want to turn the profession into something akin to weather forecasting; A never ending search for more data points and more computational power. Then taking credit when a forecast is right but blaming the data when you’re wrong instead of admitting it’s all just an educated guess with more than a little insight gained from experience.

 

[Fernando]

Basically, he (and many others) want to turn the profession into something akin to weather forecasting; A never ending search for more data points and more computational power. Then taking credit when a forecast is right but blaming the data when you’re wrong instead of admitting it’s all just an educated guess with more than a little insight gained from experience.

Need to look into input variables into their model.
Bert said MARS doesn't consider condition. How can you not consider condition when buyers have idea what kind of condition home they want.
MARS for appraisal purposes is useless if not consider condition and other subjective variables.

 

[RCA]

Need to look into input variables into their model.
Bert said MARS doesn't consider condition.

True, unless you are entering C1-C6 into the MARS regression (not recommended).

How can you not consider condition when buyers have idea what kind of condition home they want.

This is mostly for the comparable sales, which of course are fait accompli.

MARS for appraisal purposes is useless if not consider condition and other subjective variables.

That is not correct, as the condition and any other variables not fed into MARS input simply wind up in the residual. However, under the RCA method, the residual is used for extracting value for these kinds of variables. The beauty is that the appraiser has a lot of flexibility in assigning such values under the constraint that they add up to the residual.

The generated MARS model won't have any contributions for variables that are not entered into it. In my case, I do not use condition C1-C6 or quality ratings Q1-Q6 since I consider them nearly useless. Every variable that is not entered into MARS regression does not wind up in the regression model. However, the other value contributions are accounted for in the difference between the model estimate and the net sale prices - the "residual." Now, assuming all data error is 0, the residual is precisely equal to the value contribution of all other variables not entered into MARS. This residual can be broken down into component attribute contributions under the constraint that they add up precisely to the residual. Now, of course, there is some data error in the variables. It could be positive or negative, but we can assume, on average, that the sum of all data errors for a property is zero. There is not much we can do about that anyway. So, mathematically, if the appraiser makes sure that all of his residual attribute value contributions add up to the residual, it won't make any difference to the adjusted sale price for the given comp. Thus, the appraiser can set the value contributions of condition, quality, and so on - as long as they add up to the residual. If he wants to add $1000 to the condition, he must subtract $1000 from another variable. It won't make any difference to the adjusted sale price and thus won't impact the final SCA value conclusion.

 

[Fernando]

True, unless you are entering C1-C6 into the MARS regression (not recommended).

This is mostly for the comparable sales, which of course are fait accompli.


That is not correct, as the condition and any other variables not fed into MARS input simply wind up in the residual. However, under the RCA method, the residual is used for extracting value for these kinds of variables. The beauty is that the appraiser has a lot of flexibility in assigning such values under the constraint that they add up to the residual.

The generated MARS model won't have any contributions for variables that are not entered into it. In my case, I do not use condition C1-C6 or quality ratings Q1-Q6 since I consider them nearly useless. Every variable that is not entered into MARS regression does not wind up in the regression model. However, the other value contributions are accounted for in the difference between the model estimate and the net sale prices - the "residual." Now, assuming all data error is 0, the residual is precisely equal to the value contribution of all other variables not entered into MARS. This residual can be broken down into component attribute contributions under the constraint that they add up precisely to the residual. Now, of course, there is some data error in the variables. It could be positive or negative, but we can assume, on average, that the sum of all data errors for a property is zero. There is not much we can do about that anyway. So, mathematically, if the appraiser makes sure that all of his residual attribute value contributions add up to the residual, it won't make any difference to the adjusted sale price for the given comp. Thus, the appraiser can set the value contributions of condition, quality, and so on - as long as they add up to the residual. If he wants to add $1000 to the condition, he must subtract $1000 from another variable. It won't make any difference to the adjusted sale price and thus won't impact the final SCA value conclusion.

How convenient. Those variables not in the regression model end up in the residual. Then you still have to consider other variables in the residual with some data error.
Much faster and easier to show reader doing the old appraisal way with adjustments on most similar comps.

 

[djd09]

FTR, AO-16 is no longer part of the ASB Guidance.

isn't it just ao 40 or 41 now...what a joke

Pretext and Use of Code Words

An appraiser violates USPAP’s prohibition on pretext when the appraiser refers to
something other than a protected characteristic to conceal use of or reliance upon a
protected characteristic.43 The use of code words in an appraisal report can indicate
that an appraiser has engaged in disparate treatment, and pretextually referred to a
non-protected characteristic as a way to conceal the appraiser’s use of or reliance upon
a protected characteristic.

Examples of phrases that can constitute code word evidence of disparate treatment
include, but are not limited to, “ghetto,” “crime” or “crime-ridden,” “inner city,” and
“blight”; references to “shared values” or “undesirables”; concerns about “personal
safety due to ‘new people’”; or statements that an area is lacking “pride of ownership.”44
References to public assistance income and Section 8 vouchers can also have a coded
meaning.45 Whether a code word indicates discrimination depends on the context in
which it is used.

the word police...

 

[RCA]

How convenient. Those variables not in the regression model end up in the residual. Then you still have to consider other variables in the residual with some data error.
Much faster and easier to show reader doing the old appraisal way with adjustments on most similar comps.

99.99% of appraisal reports have different adjusted sales prices. Mathematically they all should be the same. Why? Well, because an adjustment is to make the comparable value contribution for a feature the same as the subject value contribution for that feature. By definition.

All value contributions have to add up to the net sale price by definition.

If you use a regression model to calculate the comparable and subject value contributions for those variables, with measurements that were fed to the regression input, then the price estimate from the model is the sum of those value contributions - which probably is not exactly equal to the net sale price. There will be a difference, and we can assume that difference is to the value contributions of all variables not entered into the regression input, plus some possible error. However, although we don't know the error for each property, the expected value of the error is 0. So, we assign that residual to the total value contribution of all other variables that didn't enter the regression. And we can create those variables as we think best and divvy up the residual between them. Again, mathematically, how we divvy up the residual between such variables won't make any difference to the adjusted sale price. All adjusted sale prices in such a scenario will be equal. The one caveat, the one partially subjective action of the appraiser, is to estimate the residual for the subject by ranking properties by residual from most negative to most positive and positioning the subject in that ranking by similarity to get a reasonable estimate of its residual.

You still need to figure out how this works. Sorry, but the bad news is that for this to work well - you need to be good at using MARS --- which is more difficult to understand and learn how to use - and even if you are smart, it will take time to learn how to deal with using MARS. ..... I do know appraisers who are capable of figuring this out. But from the looks of things, it is only the brightest. Get yourself in Mensa, and maybe you stand a chance.

American Mensa

With more than 50,000 members, American Mensa is the largest national Mensa operating under the auspices of Mensa International, Ltd.

www.us.mensa.org

 

[Fernando]

99.99% of appraisal reports have different adjusted sales prices. Mathematically they all should be the same. Why? Well, because an adjustment is to make the comparable value contribution for a feature the same as the subject value contribution for that feature. By definition.

All value contributions have to add up to the net sale price by definition.

If you use a regression model to calculate the comparable and subject value contributions for those variables, with measurements that were fed to the regression input, then the price estimate from the model is the sum of those value contributions - which probably is not exactly equal to the net sale price. There will be a difference, and we can assume that difference is to the value contributions of all variables not entered into the regression input, plus some possible error. However, although we don't know the error for each property, the expected value of the error is 0. So, we assign that residual to the total value contribution of all other variables that didn't enter the regression. And we can create those variables as we think best and divvy up the residual between them. Again, mathematically, how we divvy up the residual between such variables won't make any difference to the adjusted sale price. All adjusted sale prices in such a scenario will be equal. The one caveat, the one partially subjective action of the appraiser, is to estimate the residual for the subject by ranking properties by residual from most negative to most positive and positioning the subject in that ranking by similarity to get a reasonable estimate of its residual.

You still need to figure out how this works. Sorry, but the bad news is that for this to work well - you need to be good at using MARS --- which is more difficult to understand and learn how to use - and even if you are smart, it will take time to learn how to deal with using MARS. ..... I do know appraisers who are capable of figuring this out. But from the looks of things, it is only the brightest. Get yourself in Mensa, and maybe you stand a chance.

American Mensa

With more than 50,000 members, American Mensa is the largest national Mensa operating under the auspices of Mensa International, Ltd.

www.us.mensa.org

You're assuming in a perfect world, people have all the information and make the perfect sales price. Unfortunately, in the real world, each sales price have unique circumstances.
Appraisers have judgement in reconciliation in analyzing the comps. It's not averaging the comps as real appraisers know.
Within a market value range, Appraisers make better judgement of most likely point value than computers because we can reconciliate.
Mensa should know that.

 

[RCA]

You're assuming in a perfect world, people have all the information and make the perfect sales price. Unfortunately, in the real world, each sales price have unique circumstances.
Appraisers have judgement in reconciliation in analyzing the comps. It's not averaging the comps as real appraisers know.
Within a market value range, Appraisers make better judgement of most likely point value than computers because we can reconciliate.
Mensa should know that.

You have a very simplistic understanding of things. Data mining was engineered to work with real world data that typically has a lot of noise. Now the petri dish is somewhere where you can find some simple patterns.

Yes, each sale has unique circumstances. But isn't that so trivial?

Even if you have several sets of objects, events, or people, they share possession of specific attributes. We can find values associated with attributes that have a certain correlation with a target variable (such as Sale Price or DOM), between 0 (no correlation) and 1 ( perfect correlation). Where there are correlations, we can use data mining (e.g., MARS) to define the relationship between the attribute values and the target variable (e.g., Sale Price). This model can be used to estimate the target variable were it does not exist. This can be done entirely objectively. It can be replicated by different people if the data is shared.

We want to avoid the appraiser's so-called judgment as much as possible to reduce the probability and degree of bias as much as possible. We can then promise those involved in the outcome of our valuation that the result is as fair as humanely possible.

So, just about everything you have stated indicates that you don't have a clue about what bias is, that it is to be avoided, why it should be avoided, and so on.

I don't think you are doing yourself any favors by posting opinions like this. ....

 

[Joe Flacco]

You have a very simplistic understanding of things. Data mining was engineered to work with real world data that typically has a lot of noise. Now the petri dish is somewhere where you can find some simple patterns.

Yes, each sale has unique circumstances. But isn't that so trivial?

Even if you have several sets of objects, events, or people, they share possession of specific attributes. We can find values associated with attributes that have a certain correlation with a target variable (such as Sale Price or DOM), between 0 (no correlation) and 1 ( perfect correlation). Where there are correlations, we can use data mining (e.g., MARS) to define the relationship between the attribute values and the target variable (e.g., Sale Price). This model can be used to estimate the target variable were it does not exist. This can be done entirely objectively. It can be replicated by different people if the data is shared.

We want to avoid the appraiser's so-called judgment as much as possible to reduce the probability and degree of bias as much as possible. We can then promise those involved in the outcome of our valuation that the result is as fair as humanely possible.

So, just about everything you have stated indicates that you don't have a clue about what bias is, that it is to be avoided, why it should be avoided, and so on.

I don't think you are doing yourself any favors by posting opinions like this. ....

What you are saying is your model is better than appraisers using their judgement without market evidence for adjustments.

Can you upload a sample report using MARS?