You should decide in your mind which of your market area developments are Q3 and which are Q4 quality, then NEVER deviate again in what quality you call the homes in each respective area.
If you've done this (changing ratings) a lot, FANNIE MAE will send you a letter telling you that its a "no-no" unless:
1. You have personally inspected the property SINCE the last time you used the other rating, or
2. The property has sold SINCE the last time you used the previous rating for it.
The same exact thing applies to changing "C" condition ratings.
They track this stuff.
Now if you have to use a comp that you rated a "Q3" in a prior report in a new report where you are using "Q4" as a rating for subject, I recommend that you put a "0" on the grid and a comment in the addendum that the difference in quality is minimal instead of changing the rating you used before.
A lot of appraisers who've been doing this for decades were incorrectly taught long, long ago that this stuff is relative, and that so long you are consistent for the purposes of the report at hand, its OK to switch a Q rating or a C rating for a comp. But that is NOT what USPAP and FANNIE say.
And since about 2014 they are tracking this stuff and enforcing conformity and uniformity.
USPAP doesn't say anything directly about UAD C/Q ratings. You do have to comply with UAD requirements, as long as they don't conflict with USPAP. But, UAD ratings are just a communication and shouldn't impact your appraisal calculations: For example, if you have two comps which are each Q3 and your subject is Q4, that does not imply they should each have the same adjustment. In fact the adjustments could be quite different. One buyer pays for the Q3 and another doesn't. And, per the definition of Market Value and the Sales Comparison Approach, the Sale Price (minus concessions) for each comparable is the Holy Grail here. Your adjustments have to fit with those Comparable Sale Prices, not your imaginary view of what the difference in value is.
So, to be more exact, it is a kind of marriage between the sale price (minus concessions) and the regression model for each comparable, together with the regression model estimate for the subject and your personal-subject estimate of what the residual for the subject should be, that determines your adjustments for Condition, Quality, Functional Utiility, etc..
Adjusted Sale Price for Comprable N:
AdjustedSalePrice[N] = SalePrice[N] + regression-estimate[Subject, GLA] - regression-estimate[N, GLA] (this is the comparable adjustment for GLA)
+ regression-estimate[Subject, LotSize] - regression-estimate[N,LotSize] (this is the comparable adjustment for LotSize)
....
+ appraiser-estimate(Subject,CQA] - regression-estimate[N,CQA]. (this is in fact the sum of all comparable adjustments for subjective features)
where those CQA (Condition/Quality/Appeal) adjustments can be broken down between Condition, Quality, Functional Utility and any other subject features
any way you want (because it doesn't matter).
OK. Easier said than done. There is a lot of work in setting this up. A good model has to be drawn from the Market Area, not just the neighborhood. You may need to collect sales transactions going back N years to get a sufficient number of comps for regression, and make adjustments for date of sale. You have to create a model, generate estimates of the sale prices, take the difference to actual sales prices to get residuals, rank the residuals, score them 0.0-100.0% above below to get CQA scores of say 0.0-100.0 (or you could simplify to 0.0 - 10.0), then run a regression beteween the those CQA scores and the residuals to get a second Stage II model to add to the first regression model, to get an overall regression model, .. try to rank your subject in that list of transactions and then finally in the list of comprables to get a good estimate of what its CQA score should most likely be, then once you have that, you can actually apply the regression model to the subject features and its CQA score to get a value, or to follow the protocol of doing the SCA, calculate all value differences, feature for feature between the subject and the comparables to get Adjusted Sale Prices for the comparables, then weight and average those to get a value for the Subject.
And when that is all said and done, you are not done. You have to review and make sure that everything makes sense and is saleable to the client. Here are some rules:
1. The model cannot be overly complex. MARS segmented linear is nice. Earth with it's new feature may round the breaks between segments, but that is ok.
2. If the regression for a given feature goes up and down, you must have a reasonable explanation. E.g. the market does not pay extra for more than 3 bedrooms for a given GLA. In fact the value of the property may decline with more than 3 bedrooms, as more rooms means smaller rooms overall, or a reduction in functional utility. Again, this is in relation to a given GLA, as you are likely adjusting for GLA at the same time.
3. For the most important parts of the regression, the trend has to make sense. We don't for example expect the model to say that larger GLA is generally less valuable that smaller GLA properties.
4. We want to see adjustments that make sense. This can be a problem, because there is a lot of collinearity between many features and the regression may substitute one feature for a combination of others. When you see this, you need to generate a better model, by altering the parameters. For example, a model that adjusts garages for $150K/car, may not work for the subject neighborhood (there are some neighborhoods in California where it possibly would)
5. Finally, the client does not need to be able to replicate your model using your tools. MARS often uses random seeds and some seem to think you must always use the same random seed or keep track of the one you are using, so that you can replicate the model. That doesn't matter. You can manually create a model, as a matter of fact. What is important is that you use that model consistently across all comps, and that you can show how good your model is by using it to create estimates for all recent sales in the subject Market Area, and then look at the resulting residuals, and calculate the amount of variance accounted for by your model, that is it's R2. So, in other works, if you have a good model, I can go out and randomly find some sales comp to throw at it - and it should be able to give me an estimate of its sales price that is relatively close to its actual sale price. That is the real test of your model. You do not have to explain how exactly to recreate it.
So, yes, I can tell you all of this. But for you to go out and actually do it, is another story.
