Why is it so easy? Why is it so hard?
The upcoming joint AI and AI Canada conference is coming in June, 2017, to Ottawa. My talk this year is titled, Common Regression Errors and Mistakes.
Regression is nothing but a mathematical formula. How can it be wrong?
Not all that complicated actually – find the slope and intercept of a line, given a bunch of data points. Easy. I don’t even need to know how to minimize the square root of a bunch of squared numbers . . . just push the “regress” button, and the answer pops out. It’s just math!
Yet why does the answer seem ok sometimes – and not so ok other times? The reason: Regression is an algorithm — “a process or set of rules to be followed in calculations or other problem-solving operations, especially by a computer.” The real question: Does the regression algorithm answer the question asked? The analyst, the appraiser, must know when and how to apply regression to the problem. It is a tool. A hammer is a tool. You can cut a board in half with a hammer – but it ain’t pretty!
Well — using the regression algorithm in the wrong place, at the wrong time, will give you gobbledy-gook. Canned software, whether specialized for appraisers or not – will not tell you when to use it and how. Before you push the button, it’s helpful to know what your objective is. For valuation, there are three uses that have particular application in three parts of the appraisal problem. We must add – plots of the data come before a regression, – and work jointly to give value solutions. They are nearly inseparable. The graph communicates to our understanding, while the regression estimate or prediction provides a numerical answer.
Three basic purposes of regression:
Description is useful in understanding a market, neighborhood, or property type, or trend, or relationship between any two (or more) variables (elements of comparison). The plot helps my brain understand. The number helps me provide an adjustment or estimate or prediction.
Classification is useful in grouping or categorizing. For appraisal work, classification usually means the predicted variable is binary (yes/no), (belong/not). In particular, this type of regression can be very useful in defining what is a comp, and what is not. It helps establish the CMS (Competitive Market Segment). [In future blogs, we will consider the high importance of objectively identifying the CMS in future data-science based valuation practice.] This step is critical to replace the traditional “Trust Me” approach of picking comparables.
Prediction is useful to “support” things. This use of regression can: 1) forecast the most probable selling price; and, 2) calculate, estimate, or asymptotically converge on a reliable adjustment amount. Regression is a model. Regression is an algorithm. You don’t need to know the math. You do need to know when and how to use it. It is a model.
Regression, when used properly, is quick, effective, accurate, and convincing. In upcoming blogs, we will consider how to do it right.
Your work can be quicker, better, and not cheaper! All three!