The GP approximation 'vecchia' can currently not be used when there are grouped random effects

[mglowacki100]

Hi,

I am working with a large panel dataset and I encountered the following error when trying to initialize a GPModel: [GPBoost] [Fatal] The GP approximation 'vecchia' can currently not be used when there are grouped random effects

My setup involves both a time dimension (using gp_coords) and a categorical grouping variable (using group_data for stock tickers). Due to the size of the dataset, I need the Vecchia approximation.

My questions:

Is this limitation purely an implementation choice/roadmap item, or is there a fundamental mathematical/theoretical conflict when combining Vecchia approximations with independent grouped random effects in the same model?

Are there any recommended workarounds for large-scale panel data?

Thank you for this great library!

[fabsig]

Thank you for using GPBoost!

Yes, a Vecchia approximation is not yet available when there is a GP plus grouped random effects.

Is this limitation purely an implementation choice/roadmap item, or is there a fundamental mathematical/theoretical conflict when combining Vecchia approximations with independent grouped random effects in the same model?

Both. We might be developing something in the near future (a few months), but no guarantees.

Are there any recommended workarounds for large-scale panel data?

It depends on what you want to do. I you really have only one single, global GP shared across all data points, then I am not aware of a nice workaround (besides things such as sub-sampling). If every group (stock in your case) has a separate GP, then you can use the cluster_ids argument instead of group_data to pass the grouping data. The dimension of the GP is then the number of samples per group. This can be much faster.

Hope this helps.