The key to any statistical analysis is DATA, DATA, DATA!!
Single family real estate data is not very reliable or consistent, and not enough is available in many areas, as we all know.
CU is the most significant attempt to get more useful data by requiring appraisers do use specific coding and criteria. However, real estate is local, local, local. Even the number of bedrooms varies a lot as there are different criteria for determining what is a bedroom, even in the same city. Three appraisers measuring the same house will probably not have the same square footage, as I learned doing relocation appraisals.
With CU, this is becoming more obvious as there are sometimes wide variations in how appraisers code factors. For example, why do condition ratings vary? How accurate is MLS? Is public records accurate? What is the best source?
Now that regression software is popular with appraisers for getting adjustments, I have been thinking about why it is often not very reliable. To understand even simple regression requires knowledge.
My first statistics class was in 1963. The first time I used multiple regression was in graduate business school in 1979, when I did a mini-thesis on factors in REIT stock volatility using SPSS.I used a remote university mainframe that kept blowing up and erasing my data. There were no data issues. Doing multiple regression analysis on real estate housing data was not possible. Way too much lack of usable data.
Since I started my Appraisal Today newsletter in 1992, I have been writing about AVMs. The less data that is available, the less reliable the value.
As we all know, AVMs work well in a conforming home in a large tract of similar homes, built in the past 10 years. After that, the accuracy and reliability goes down fast. Just check what Zillow’s Zestimate against your appraised value.