Appraisal Statistics: Regression - Do Your Really Know How Much Work It Is?

[Joe Flacco]

IF the ONLY difference in the sales was Interior vs EndUnit, that is to say, all Interior units sold, market conditions adjusted, for X dollars and all end units sold for Y dollars, then you don't need to go any further. But what if there are other differences? You will need then to extend your data set and things start to get complicated, and then you would be advised to use a regression tool.

Computer will never consider all of the differences. Is all C3 condition the same? Is all Q3 quality the same? The range in one rating is large and in reality most properties probably fall into a borderline of ratings. Computer would just assume C3 is C3 or Q3 is Q3. It is just better to minimize the differences as much as possible and do a paired sale. You can't explain away real estate with 6 Q and C ratings and a hand full of other data points.

 

[RCA]

Of course, I agree totally. I don't like the ratings. I would prefer 0-100 percent ratings; which would be X% of the property in the subject neighborhood is in worse condition. BUT, then the problem is that your number only relates to your definition of the neighborhood. A more acceptable choice would be X% of the property in the subject census tract is in worse condition. THAT would work -- but you really have to have some idea about the range of conditions in the subject census tract. But, you know, I would be a lot more comfortable with that. Why - because a lot of neighborhoods are almost entirely C-4, or Q-4; yet buyers can still differentiate condition differences that result in price differences. Now, you will likely mention that rating this way is subjective. Yep, no way around that. That's why "geographical competence" is important. If you are too far off in your subjective rating, it can be caught by the "data collector" in this way: They can go through all of your appraisals and see the range of condition percentages you have been giving out. They should be EVENLY distributed between 0 and 100%. If all of your condition percentages are over 50%, you are in trouble. Or they could just analyze your appraisals that they have on hand. Or they could compare to the other appraisers. You are long term screwed, if you don't try to do a good job on your percentage scores for condition and quality. But the people who decided this, were a bit too timid, - just a bit too timid. The end result is that for all practical purposes, we don't have a lot to go on for condition and quality. But it is about all we have.

 

[Joe Flacco]

What impacts value in one neighborhood may not impact value in another neighborhood. How value is impacted in one neighborhood or subdivision may be impacted differently in another neighborhood or subdivision. Big data can't be more reliable than smaller sets with very similar properties for real estate.

 

[Joe Flacco]

I'm no expert on regression or statistics. I am just giving my point of view based on my observations of the market.

 

[Ariba]

How good is a regression model in providing the adjustment for style: bi-level, tri-level, 2-story; view, utility lines, solar, radiant heat, etc . Also, the regression model uses raw MLS data, which in many instances the information is wrong.or not available Real Estate agents don't even know the difference between a 2-story and tri-level. Are the bedroom conforming or non-conforming? Interior or walkout basement. Regression model is a great academic exercise. Bottom line, if the regression model is so great why do AMC or lenders need appraisers?

 

[RCA]

What impacts value in one neighborhood may not impact value in another neighborhood. How value is impacted in one neighborhood or subdivision may be impacted differently in another neighborhood or subdivision. Big data can't be more reliable than smaller sets with very similar properties for real estate.

I'd like to say that, generally speaking, you are correct. Getting a high R2 on large data sets that span different market areas is more difficult than data sets that are within a given market area. But I would be cautious. Just think, when you build a data set, it is on past data, i.e. past sales. It is always limited in that respect. There is never any guarantee that when you bring in a property under new market conditions and maybe with new features, it will get a reliable prediction. Thus one could argue that expanding your data set will make it more robust in predicting new kinds of property sales (or virtual sales if doing refinancing). But generally, rule of thumb, you want to try to narrow down your set of comparables as much as you can to be similar to the subject property. And if it is relatively similar to the comps you are OK. But, if has one or more important features not in the comps, like an astounding view of the ocean, immaculate remodeling, then all bets are off - but of course you said "similar".

 

[Michael P Jacobs MAI]

<SNIP>In other words, by using less sophisticated techniques, the sales may adjust to a range between $815k and $835k; the appraiser picks their point within that range based on the rationale described in the reconciliation. Assuming that range is reasonable and credible, I'm trying to gauge how material it would be to go deeper and use a more sophisticated technique?

Bingo! But I would like somebody to be able to argue this *exactly* in deep-stats-language because as practical as you've phrased your question isn't intellectually compelling enough for the wonks that are more in love with the method than the outcome. Your best argument would include something along the lines of (1) you need an inspection anyway (supported by math), (2) if Boots the Appraiser isn't going to be able to determine a precise estimate of value Big Data is likely to be off even more (show more math here), and (3) Boots is more cost effective and consistently reliable (finish by applying Monte Carlo analysis to variants on your prior math).

 

[Michael P Jacobs MAI]

<SNIP>... Understand this is not really exact, as the initial classification can be quite arbitrary...<SNIP> ... The statisticians who developed CART and MARS are concerned with the issues you mentioned, because they affect the outcome and they have developed fairly reasonable techniques to overcome them. In fact, refer to the link below on CART and you will find the term “heteroskedasticity” mentioned eleven times, with a section on it...

"Understanding this is not really exact..." sounds a lot like the appraisal process but I appreciate the way you try to simplify the process. The software naturally has come a long way since I studied this years ago - I feel like I'm trying to understand a modern programming language from a punch-card perspective!

I briefly read the heteroskedasticity section and felt myself being pulled again by the dark side and, after all these years blissfully honing my human intelligence at the expense of keeping up with artificial intelligence, I'm going to have to walk away from this discussion to make some more money. There's a drug to this topic, and as a wonk myself I can feel the serotonin brain-bath dulling my perspective. "How'd your day go, hon?" Well, I didn't make any money but I really enjoyed getting a glympse of MLS data in four dimensions again..."

Good luck, Bert! I'm more interested in the policies and safeguards of AI in real estate valuation than the application these days. I'll leave you with the thought that just because we "can" doesn't mean that we "should" from an economical perspective. I hope in your continued R&D you do not stop arguing for Mr. Boots, the appraiser-on-the-ground to be part of the process. It would be a shame to "prune" the human intelligence and find that it really was the best way after all.

 

[Joe Flacco]

I'm of the opinion residential appraisers should avoid trying to mimick avms altogether and be better at the basics.

 

[BRCJR]

If you use regression analysis in your report are you not required to be able to explain how it works, as in understand the method/algorithm employed?