I'm trying to model the likelihood a customer will be delinquent on their loan by next month based on their most current data at the time. Currently, I have longitudinal data and am having some trouble understanding the correct way to split my data into training and testing to get a more accurate representation of model performance. So the 2 approaches I could apply are:
This could bias results as customer's behavior may be similar in adjacent months and cannot be considered independent. So for eg: If I split such that Customer X at month 6 is in training and Customer X at month 7 is in test, the model will likely be able to predict correctly as it's learnt similar data from month 6.
Form Groups of Customers and split
This approach seems to be more independent of customer since all of a customer's data will be either in training or testing, but not both.
Wanted to know if the second option sounds like a better way of modeling this behavior. Would appreciate any other splitting suggestions I may not have thought of.