Tracking of Model Performance Over Time in Logistic Regression

Now, suppose you conclude, based on the previously mentioned criteria, that the model is not stable. What will you do then? Let's hear from Hindol on that.

Let's go back to the telecom churn example. If you recall, the model was built using data from 2014. Now suppose, you are tracking its performance over time, and that ends up giving you the following results:

So, the first time, when the model's Gini dropped to 0.72, you avoided building a new model. Basically, you just recalibrated, i.e. updated the coefficients of the variables. That resulted in a slight increase of Gini. However, the next time Gini dropped to a low value, i.e. 0.71, we just rebuilt the model, i.e. got new sample data, performed data prep, etc. and built the entire model.