We should create a dedicated "error" model to predict either the % error or absolute error from the main model. This would be useful for a number of reasons; it would:
Help us look for incorrectly validated sales. If the predicted error vs actual error are extremely far apart, then we could be dealing with a non-arms-length sale
Let us know where to prioritize for desk review, based on areas with high predicted error
Target individual properties for desk review, as high error may indicate characteristic errors, etc.
We should create a dedicated "error" model to predict either the % error or absolute error from the main model. This would be useful for a number of reasons; it would: