Metric for model selection in bayesian time series modelling(focusing on causality, not prediction)

To give some intuition behind the task.

Assume we have several different parametrizations of the same causal graph, one parametrization can e.g include interaction/moderation effects and one just main effects.

We are working with time-series data.

The goal is to discern the causal effect of our predictors, even though prediction is also important in the task the main goal is to discern the causal effect.

I commonly see LOO-cv with the psis-loo algorithm being recommended as well as WAIC as metrics for model selection tasks. I however question their validity for time-series since we will take into account future observations for our predictions today…

Has anyone read any literature with regards to this task or can provide some pointers on how to go on about this?