We collect administrative fees from our customers based on many complex business rules albeit based on few variables. I have the history of fees colected through time (about 500 records for each customer, for over 10k customers) as well as the the variables used in the calculation, such as his number of customers, total revenue, and date of the payment.

I am attempting to build a model that, for a given payment date and other variables, would estimate the amout to be collected in administrative fees at that point in time.

As payments are space and varies in dates I imagined something more sofisticated than a LM would berequired. I am trying to adapt this script to my data, though would require me to generate one model for each customer and I wouldnt feel confident It can be trained with such little data.

Would you please advise on how to better model the use case? Went with Transformers and Time Embeding experimenting. Please feel free to point out other algorithms that would eventualy do a better job.

  • $\begingroup$ Welcome to DataScienceSE. Sorry but if the company knows how to calculate these fees precisely in a deterministic way, what is the point of training a complex model which would predict the fees less accurately and less efficiently? $\endgroup$
    – Erwan
    Nov 22, 2021 at 23:29
  • $\begingroup$ hey thanks! the goal is not quite to calculate the fees themselves, but rather to detect anything different from the expected. E.g. upon receiving a new payment I'd check if too different from the predicted and flag it for auditing in case its too far off. I could use such model to guide where to put auditing efforts. $\endgroup$
    – filippo
    Nov 23, 2021 at 2:24


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.