This PR removes property class as a predictor and adds Census tracts.
Class is dropped due to concerns about its influence (it's highly predictive) compared to its accuracy. If it's included in the model but a property has an un-updated or otherwise incorrect class code, then they may receive an incorrect value. Class is also basically a combination of features already in the model (sqft, age, etc.), so removing it should not substantially impact performance.
Adding Census tracts is just one additional way to further split up location.
dfsnow
commented
5 months ago
This PR removes property class as a predictor and adds Census tracts.
Class is dropped due to concerns about its influence (it's highly predictive) compared to its accuracy. If it's included in the model but a property has an un-updated or otherwise incorrect class code, then they may receive an incorrect value. Class is also basically a combination of features already in the model (sqft, age, etc.), so removing it should not substantially impact performance.
Adding Census tracts is just one additional way to further split up location.