Tuesday, July 9, 2019

Why Major Mass Appraisal Jurisdictions should Hire AI Engineers

“AI engineers don’t write code to build scalable data pipelines like a data engineer...instead, they understand how to extract data efficiently from a variety of sources, build and test their own machine learning models, and deploy those models using either embedded code or API calls to create AI-infused applications."

Conversely, Regression models are not intelligent as they are highly modeler-dependent (subjective). Thus, given the same sales dataset, five modelers may come up with five different models with very different results. Of course, the biggest failure is the Sales GIS (very dynamic) as it's representativeness to the population (more or less static) is difficult to establish. That is why, in my AVM book I proposed and reverted to fixed neighborhoods (not sales dependent and are population-derived).

AI engineers do not use any data variable modeling. Their data extraction process is extremely smart leading to very smart machine learning models. In mass appraisal environment, they will be able to precisely identify and demonstrate where the sales datasets and populations are at variance. Whereas, the mass appraisal (known as “cama”) modelers will simply remove them from the model as outliers, creating unexplainable gaps and major fault lines they won't even know.

In our future consumer environment (Homequant, Homeyada, Condoyada, etc.), we may ask the first-time users (optionally) to take a three minute tutorial. As the user interacts with the tutorial, our machine learning models will extract the data (by reading the responses) and fine-tune the model, specifically for that user. When the user returns to value a subject (log-in will be needed to identify the user), our model will populate the comps as soon as the subject is defined. So the ten minute exercise will be reduced to fifteen seconds.

Alternatively, in a traditional modeled environment, it's all sample-based so the results are, at best, bell-curved with the customary 68% efficiency. The error-based cama regression models fail to test the optimality of the solution; for example, is a model Coefficient of Dispersion (COD) of 8 better than a COD of 10? The COD of 8 could represent a post-optimal solution, whereas the COD of 10 could perfectly represent an optimal solution. But in cama environments, the COD of 8 would be universally preferred. In fact, several years ago, I presented a paper along these lines at a conference, raising some serious questions.

The mass appraisal industry is extremely antiquated. They are still using the 30-year old Regression Modeling and mostly Sales GIS. That is why I think the major mass appraisal jurisdictions should hire AI engineers, proving that the industry needs to look ahead. 

Granted, given the paucity of AI engineers, it is not going to be easy but they should try. They should also remember what Steve Jobs said, "It does not make sense to hire smart people and then tell them what to do. We hire smart people to tell us what to do."


Thanks,

Sid Som MBA, MIM
President, Homequant, Inc.

No comments:

Post a Comment