We are working on a project to forecast direct deposit amounts for bank members using GluonTS DeepAR.

Here’s the setup:

• We have 3 years of direct deposit data for multiple members.

• The deposit frequency varies:

     1. Some members receive deposits weekly,
     2. Others bi-weekly, and
     3. Some monthly.
  

• To create continuous time series for all members, we:

     1. Added missing days to ensure a consistent daily frequency.
     2. Imputed the missing transaction amounts with 0.
  

• As a result, most time series now contain a large number of zeros, with only occasional non-zero deposits.

The Challenge
When using DeepAR from GluonTS, we are:

• Getting either negative or zero amount forecasts, which makes sense given the dominance of zeros.
• However, this leads to poor accuracy, as the model struggles to predict the occasional deposit amounts correctly.

Questions

• Is GluonTS DeepAR suited for this kind of sparse time series forecasting with many zeros?
• Are there any techniques or distribution settings in GluonTS to handle this better?
• Any other modeling strategies or tips for improving accuracy with DeepAR on sparse time series?

Would appreciate any insights, suggestions, or references to similar projects.