README doesn't cover how the simulator decides which transactions are fraudulent

[richsanchezalexander]

Two follow-ups, both about the generator rather than the data itself:

1. Is there a description anywhere of how the simulator decides which
   transactions are fraudulent? The README doesn't cover it and I
   couldn't find it in the ICASSP paper.

2. Related: online transactions ("ONLINE" merchant city) are 40x more
   likely to be fraud in 2015-16, but 5x LESS likely in 2018-19. That
   reversal is large enough that a model trained on the earlier period
   learns an inverted rule. Did the generation logic change between
   those periods?