I have a question regarding creating a transactions dataset from scratch.
I've created customer profiles and am generating transactions based on these profiles. The way I do this is based on the method described here: https://fraud-detection-handbook.github.io/fraud-detection-handbook/Chapter_3_GettingStarted/SimulatedDataset.html#fraud-scenarios-generation
The labelling will be dependent on credit card credentials + user preferences. I am currently labelling transactions on rules like this.
p% of the time a customer's purchase amount is either in the 95th percentile or 5 percentile of the mean amount of their purchase will be marked as fraudulent. (p is any percentage)
I've made many rules like this based on intuition and common fraud signals. Is there a standard or accepted method to assign transactions a label when creating a dataset from scratch that will yield accurate results for anomaly detection models like AWS Fraud Detector or would this kind of ruling system work?