Regression for GLA Adjustment

[Mejappz]

"Credibility" has to be the lowest bar among all standards ever devised by humans, and is essentially untested in the appraisal world because no one will pay to compare the result to reality. Anyone can write a credible report completely lacking in veracity, and since the practical application of "credible" has become, "can we make the loan?", there really is no standard in that. You simply cannot reconcile the "most probable" aspect in the definition of market value with the lack of complete lack of value in "credible" results as "supported", our other non-existent standard of quality. The clear result is that the bulk of lending valuations will soon be conducted outside the toothless/useless/worthless auspices of USPAP and the existing licensing and regulatory bureaucracy.

Yep, I couldn't agree more.

 

[RCA]

"Credibility" has to be the lowest bar among all standards ever devised by humans, and is essentially untested in the appraisal world because no one will pay to compare the result to reality. Anyone can write a credible report completely lacking in veracity, and since the practical application of "credible" has become, "can we make the loan?", there really is no standard in that. You simply cannot reconcile the "most probable" aspect in the definition of market value with the lack of complete lack of value in "credible" results as "supported", our other non-existent standard of quality. The clear result is that the bulk of lending valuations will soon be conducted outside the toothless/useless/worthless auspices of USPAP and the existing licensing and regulatory bureaucracy.

I disagree. However, there are a couple of problems. You could compare value conclusions to subsequent sale prices adjusting for trends over time, and we would expect some correlation if the value conclusion indeed accurately reflects market value on the effective date of the appraisal. But the actions of specific buyers usually do not reflect the most probable actions of a future buyer in some market area. They are really just random instances of activity between buyers and sellers in the market. For a specific property would we cannot expect to see more than one such transaction every 8 years or so. In 8 years so much can change, including the condition and updates on a house, that we can no longer make meaningful comparisons. So, while this might be an interesting statistic, it will not really do the job of assessing accuracy. All we are left with is assessing how objectively and correctly an appraiser does his job, with 100% accurate SOW and Data inputs. We have to look at his process at a microscopic level and assess what could go wrong. This latter technique kicks the traditional SCA out of the ballpark - in a bad way.

Now, "making the loan" of course has absolutely nothing to do with credibility - except from the perspective of a lender or sales agent who's commission, promotion or sadly - job - is on the line in getting a transaction to successfully complete. As appraisers we back up - and look at this from a 0 - 12 +/- year period after the transaction closes and consider what is going to be lost by someone when the loan forecloses resulting in sale of the home. It will be sold at liquidation value - which we assume will bear some relationship to the appraised value. One can certainly question the logic of this. Homes are often trashed in the process and all kinds of auxiliary things can happen that will impact any real loss.

But look at this more generally as the valuation of anything, where this valuations impacts decisions, which in turn impact the efficiency, profitability and survival of people, corporations and the government. It is just a cog in the wheel - but an important one. Valuation in general is a problem in itself. That needs to be solved. Real Estate valuation is just a special case - although one of the largest categories of assets on the planet.

However, the problem is that people are looking for a way to compare a standard appraisal done without regression, utilizing adjustments pulled out of thin air based on so-called experience, with one that is done according to a strict protocol based on the objective use of a regression tool that precludes selection of adjustments based on nothing more than subjective experience. Once you accept this per-condition, then you might as well stop in your tracks. We can't look inside the head of an appraiser to see what he is thinking and determine why he decides this or that.

Credibility comes in several stages. The appraiser only has limited control over the quality of certain stages. We can however discuss how believable the appraisers analysis and valuation is, under the assumption that all sources he relies on for data and scope of work are perfect.

1. Scope of Work: The appraiser can be hemmed in by constraints imposed by the client on Scope of Work. Typically these impose restrictions on time for inspection, analysis and preparation of work. Of course the appraiser is expected to reject an SOW if he feels it would impact the credibility of the appraisal. So, let's make this easy. Let's assume there are no issues with the SOW, whether or not the appraiser decides to do the appraisal.

1. Data: The appraiser gets most of his comparable data from some MLS and property records drawn from county data. He should use reasonable effort to judge the approximate quality of the data and make corrections where feasible. This can be tedious work that involves, in many cases, making an educated guess or approximate measurements to fill in missing pieces of information - which often is better than not doing anything at all. When all is said and done, we can usually assume that some at least small percentage of data is still not accurate.

But for the sake of measuring the appraiser's "accuracy", we should assume, given reasonable actions to remedy data problems, that the appraiser cannot improve the situation any further and therefore assume the data at this point is 100% correct. Yes, one can make many arguments here - and protocols for dealing with data problems can and should be established. But, let's move on.

2. Regression Analysis. The regression tool needs to be capable of obtaining maximum R2 values without over-fitting. It should be able to discover those variables that account for the greatest price variation. It should be able to deal with range changes in incremental value contribution. It should at the same time present a model that is capable of human understandable explanation. MARS is currently the only regression technique that satisfies these requirements. The choice and utilization of a regression tool is in the domain of the appraiser. The regression, its results and model interpretation, if flawless in technique, can be assumed to provide 100% accuracy and be completely objective up to the assignment of a residual score to the subject.

3. Assignment of a Residual Score to the Subject Property: This is critical and assumes the regression has been done flawlessly and cannot be improved, It also does assume that the subject market data makes sense (and a few markets don't because of a lack of activity or poor data). The appraiser has to rank subject against the set of comparables based on their residuals from running the model on their features and taking the difference of its estimate and their net sale price (net without concessions). The ranking of the comparables will typically be based on their unmeasured or subjective features such as condition, quality, functional utility and so on. This is often scored 0.00-10.00 and called CQA for Condition-Quality-Appeal. You will typically see Fixers at the bottom of the with CQA scores near 0.0 and the nicest and most Luxury homes at the top with scores near 10.0. So, the ranking of the subject property is typically not that difficult. Just find the most similar homes that are better and worse and slide it between them. In fact, in the critical areas with the steepest price curves at the bottom and top of the ranking, you can typically do a very accurate rating of the subject, so that this is method is particularly accurate in these extremes - compared to the tradition SCA approach.

Within the above context, I can kick out accuracy of the RCA approach I use in many cases. It will depend on the subject property, its market area, the data sources and it will assume I have control of my SOW. However, if I do not believe that I can create a credible appraisal, and accurate value conclusion, I will reject the assignment.

 

[sputnam]

While I agree that an appraisal is an opinion, regression can test how credible and accurate that opinion is. In the analysis below, all 46 sales were within 5% of their actual price. How well do you think it predicts the subject? Look at record 47. It is estimated that the subject would sell for 570K. It sold for 565K. There is a 95% chance the subject value is between $541,000 and $598,000.

And when was the last time your Client was happy with an appraisal that indicated a value of between $541,000 and $598,000? As I have clearly stated, regression is a tool. It's helpful.... sometimes.... depending on the quality and quantity of the data available. In some markets, you would struggle to find 47 residential sales... much less 47 sales with some similarity to the subject property. Appraisers should have an entire bag full of analysis tools... not just a few.

 

[J Grant]

While I agree that an appraisal is an opinion, regression can test how credible and accurate that opinion is. In the analysis below, all 46 sales were within 5% of their actual price. How well do you think it predicts the subject? Look at record 47. It is estimated that the subject would sell for 570K. It sold for 565K. There is a 95% chance the subject value is between $541,000 and $598,000.

View attachment 77398

Too bad our appraisal assignment is not to predict a sale price.

The above is all about sale prices.
We are developing market value opinion for a subject. We are not tasked with guessing, predicting, estimating, or opining a present or future sale price of the subject. The above runs a regression, of 47 sales, likely many of them are not comps for the subject to get a range. Then what happens...aprice is picked from the range....and it gets called a value? That makes it a price opinion labeled as a value opinion. But none of the tests of market value have been applied. None of the adjustments, if any were even made, are known. Other than a bunch of prices, what evidence is there that these 47 sales are similar to the subject? If only 10 out of the 47 are truly similar/comps that means data from 37 not comparable sales are skewing results.

Even in a relocation or listing use appraisal, we still are developing an opinion of the market value for the subject. The expectation is that it align with a future price, but it is still (typically) a market value opinion. An appraisal is defined as a value opinion. Can anyone show me where appraisal is defined in USPAP as an opinion of price? ( rhetorical question/sarcasm intended )

 

[Zephyr999]

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.

Something is wrong with one assumption or the other...if 'experience' is based on same data set.

 

[RCA]

And when was the last time your Client was happy with an appraisal that indicated a value of between $541,000 and $598,000? As I have clearly stated, regression is a tool. It's helpful.... sometimes.... depending on the quality and quantity of the data available. In some markets, you would struggle to find 47 residential sales... much less 47 sales with some similarity to the subject property. Appraisers should have an entire bag full of analysis tools... not just a few.

You need to have established protocols that define exactly how the value conclusion is reached through objective means. "A bag full of analysis tools ..." implies a corresponding bag full of protocols -- which you will need to learn. Just coming to some value conclusion based on X, Y and Z, where X, Y and Z are themselves not based on specific recipes or protocols, is not worth that much.

The traditional way of doing SCA does not provide an explicit recipe or protocol that covers every step - in particular how to deal with unmeasured attributes such as condition and quality of construction. It is pretty much worthless.

- And any suggested protocol should be mathematically sound.

 

[Mejappz]

And when was the last time your Client was happy with an appraisal that indicated a value of between $541,000 and $598,000?

The range after adjustments was $566,500 to $584,500. If the client wants a tighter range from the Regression model(from the 46 sales) you would simply lower the confidence intervals from 95% to 90% or 80% etc.

 

[NJAPRAZE]

The range after adjustments was $566,500 to $584,500. If the client wants a tighter range from the Regression model(from the 46 sales) you would simply lower the confidence intervals from 95% to 90% or 80% etc.
View attachment 77408

so the data is only as good as its manipulated?

 

[Terrel L. Shields]

so the data is only as good as its manipulated?

The range remains the same. Within that range is likely to be the number that is "true" but since we cannot know it to be true, we can use our judgment to determine a 'final' value. But without clues (one sale is better than the other, or one requires fewer adjustments) we are equally confident (95%) the answer lies within that range.

 

[George Hatch]

In real life
The buyers and sellers in the SFR markets don't do any number crunching, and neither do the brokers who list and sell the properties to the buyers and sellers who aren't doing any number crunching or adjustments. The sellers look at a bunch of sales to see where their property fits in, and the buyers look at a bunch of sales in their price range and rank them.

That's the valuation model they use, which leads to the sales transactions we use.