Classification with dynamic features (fraud detection) - Any resources?

Let's say I am dealing with a fraud detection problem classification problem.

Each observation is a contract with static features such as length of contract, paid amount etc.
But it could also have dynamic features such as days since contract signed, days left for expiration etc.

How do i deal with those as they evolve over time? In the no fraud cases for example there will be no observation of no fraud for e.g. days left for expiration =20, 30, etc.

There is also the matter of independence of observation for most conventional machine learning techniques.

Such kind of model would also need to be able to provide predictions at different timestamps during the evolution of a contract.

I would appreciate it if someone could provide some useful resources for such problems (if there is anything available in tidymodels even better).


Conventional thinking might be the mutate function combined with the lag function to construct new features, treating each point in time as an observation, if you want to deal with dynamic features perhaps using deep learning tools such as LSTM, informer, etc. would be a good way to think about it, and you can use the keras api to construct them from R calls in python, or use the native Rtorch package