Gradient Boosting

[RCA]

For those who strive to become more erudite in the field of modern statistics, I came across a very good intro to Gradient Boosting that I am SURE you will understand: http://blog.kaggle.com/2017/01/23/a-kaggle-master-explains-gradient-boosting/

In short, what this kind of does is predict something like sales price (the so-called "target" variable), subtract the prediction from the actual to get a residual and then try to predict the residual from the input variables again, and again and again …. until the residuals get too small to care about.

Now, I use something I call the "Regression Residual" method … which is, to be very clear, not Gradient Boosting. But Gradient Boosting is used in mass appraisal and AVMs, so it is something good to know. And, I could conceivably use it … but I don't think it would make a lot of sense for a residential appraiser.

Next you might check out the post on Random Forests: https://gormanalysis.com/random-forest-from-top-to-bottom/.

 

[KHS445]

Not knocking the theory, but even the author of the article didn't completely know what he was talking about. At least two commenters pointed out errors in his article. Don't get me wrong statistics have a place, but accuracy increases exponentially when you can remove the non-quantifiable factors. In large market areas where there are sufficient comparables to remove most of the non-quantifiable items in your selection process this approach may work. However, in rural and most of the flyover areas of the Country there are not enough true comparables that all adjustments can be reduced to a formula. It still takes strong knowledge of the area and communications with individuals involved in the transaction to understand at least some of the buyer/seller motivation.