Should I use "Recency" as an predictor for churn if I want to catch churners early?

Question

I want to build a customer churn prediction model that predicts probability of churn the next day and I'm looking for some features that might be important for the target variable which has outcomes churn (1) or not churn (0).

A customer is considered to have churned if more than 365 days have passed since the date of last purchase. Naturally then, "Recency" (Time since last purchase) will be an important predictor for predicting churn. So if a customer is on his/her 364th day of purchase-inactivity the model will with high probability predict a churn next day. But I want to be able to predict churn sooner and not depend too much on time since last purchase.

How would I go about if I want to be able to catch churners early and not depend on time too much?

Answer 1

You may calculate the ratio of churn-90 customers who became churn-365 to understand the problem deeper. I expect it depends on industry and other factors. In general, churn-a to churn-b ratio may give you a better vision. Then plot different combinations of a and b. The last step is to pick a reasonable threshold that definitely involves some unpleasant trade-off between the model usefulness and model accuracy