RCA
Elite Member
- Joined
- Jun 27, 2017
- Professional Status
- Certified General Appraiser
- State
- California
Maybe I'm overly optimistic
But I've run into a decent number who know how to code, actually know statistics, etc. George Dell (for example) has had a number of appraisers go through his course.
I have never met an appraiser who has used Salford Systems MARS (now called Minilab Salford Predictive Modeler) or R/earth for real appraisals. I have talked to Dell and read some of his writings, but to date, I have no indication that he is using MARS or R/earth for a full appraisal. I am sure he has tried it. But it is more than just being able to run earth() on some data.
So, yes, - it is one thing just to run earth and get a MARS model, and another thing to use R code to take that complex model and turn it into a Sales Grid.
I have had people like DW tell me it is all just "high school algebra." Of course, they have passed a number of college courses in Calculus, Linear Algebra, and the like. It is one thing to pass a college-level course in math - and another thing to wrestle through the complications to apply perhaps simple math concepts. You take multi-stage algebraic manipulations cleverly, switching all kinds of things around, staying within the algebra, keeping it all together in your head to achieve some elegant end, and hope not to forget some of the necessary links in the process. Then, ask yourself if you know what you did. Sometimes, it is applying one algebraic transformation on top of another on top of another, and so on. '
You could take someone with a college degree in math and give them a MARS model - "simple algebra" - and then ask them to give you the formula for the value contribution of additional GLA for a 2500sf home or from the model calculate the GLA adjustment between a 2000sf and a 2400sf home. They will likely scratch their head if they have never dealt with such problems before.
I also see this with Claude, for example. I ask it to code something to save me some time and mental strain - and it often does surprisingly well (not always). If I keep going with Claude to add more things, I can be pretty sure that Claude (or ChatGPT) will eventually break down and turn what was once good into a total mess. And then I take some time to look at the mess and tell Claude: Why didn't you do it this way? -- And it comes back and tells me that my solution is elegant and then immediately starts spitting out good code again --- at least for a while. There is another problem: Claude will eventually say I have run out of tokens for the day, and I have to wait until 2:00 AM to continue. But it appears to have lost part of its memory when I continue. Just like people!!! It can only hold so many tokens; it only has so much memory, and it eventually forgets things. You could try to reboot it with the last code it generated - which is usually a history of snippets - but you will run out of tokens again, - sooner rather than later. So, people are like that.
Of course, LLMs like Claude are really not "smart"; they are neural networks, trial and error prediction, incrementally getting better, and so on. You could say the same for some people of course.
Most people can handle only so much algebra - and in reality, the amount of algebra contortions needed - is often more than they can handle.
So, when someone says, "It is only high school algebra, " they are showing their lack of knowledge. One could maintain that mathematics is only algebra. More on that at a later date.
I'm eager to hear from those who have tried using MARS (R/earth) for appraisal and those who feel they have succeeded.
To "succeed," you need to be able to get, in most cases, an R2 of over 70% (0.70+) for several appraisals, with a CVR2 (cross-validated R2) of about 60% or higher and then apply that model to the SCA approach to value. I started doing that in about 2004 without using RCA constraints, which is nonetheless quite satisfactory and, in any case, much better than using linear regression. You do have to learn to detect over-fitting and correct it, if possible (with small sample sizes, you may have to live it, although with a small sample size, you can only assume it is overfitted.) You also must have acquired a pretty good understanding of the subject market area, which using MARS helps.
