So I’m working on a monitoring model and could use some input on what sort of statistical method would be suitable for this particular case.
Imagine you are a financial institution with corporate customers transferring and receiving money to/from countries all over the world. You want to be alerted when customers diverge from their typical/expected behavior. One simple measure could be monthly aggregated transactions where the counter party is located in a tax haven. By “typical/expected behavior” we are thinking in terms of the industry segment the customer belongs to, e.g. the behavior of a retail company is likely to be very different from that of a shipping company.
One thing we are on the lookout for is when a customer “suddenly” has a big increase in transactions with such countries. The absolute amounts are often not as critical as the relative increase (e.g. a customer has a monthly amount of 5x its monthly average for the last three months). But one source of false positives here, when we compare the current month to immediately preceding months, is a “seasonality effect”. Perhaps it is common for one industry segment to have substantially larger transaction amounts in the summer months, another one just before Christmas, etc.
I’m looking for a smart way to figure out threshold values here. For example: If I am running my monitoring model on 1 April 2023, looking at aggregated transactions per customer for March 2023, what should be the relative increase (compared to say the average of the preceding three months) allowed for customers of different segments without triggering an alert? Imagining data access is not an issue (years of transaction data are available), any suggestions on what methods might be most suitable to look into here?
