Splitting Fraud Data Without Breaking the Ratio (Day 3)

Day 1 was setup. Day 2 was looking at the data. Day 3 was the first real modeling decision: how do I split this data without breaking it? The problem it solves You can't train a model on all your data