How to sample efficiently from time series data?

[junpenglao]

@light_kira

However, this creates one Metropolis sampler for each B[t] and does not scale well for a long time series. I would like to sample the entire B at once with one sampler.

You need to write the for loop into a theano.scan for performance.

@light_kira

I also have a transition distribution for B in mind that might help the convergence.

You can specify that in https://docs.pymc.io/api/inference.html?highlight=metropolis#module-pymc3.step_methods.metropolis with proposal_dist kwarg