We use Variational Bayes to recover causal structures and model human learning in a context where variables are represented by continuous-times series. This approach allows us to model learning about hyperparameters which have been assumed to be known by participants in previous research. We test two factorisations for the approximate posterior: (1) a normative one, which keeps the true joint distribution and (2) a mean-field approximation. Variational models remains effective at recovering the correct causal structure and also manage to estimate hyperparameters. Both factorisations reduce the model's ability to disentangle direct and indirect effects, a well-documented human error which we here replicate with 420 participants. This initial agreement of Variational Bayes with participants' behaviour encourages further modelling attempt, specifically those considering how interventions shape learning under its constraints.