Sampling from GaussianRandomWalk

[sachinruk]

It seems that you cannot sample from a GaussianRandomWalk looking at https://github.com/pymc-devs/pymc3/blob/master/pymc3/distributions/timeseries.py#L134 or any other distribution in timeseries.py. Think this is partly due to the fact that the initial (default) distribution is defined to be Flat.

Just wondering if it's worth changing into something like uniform(-2,2) for the sake of being able to sample. If that is the case, I'm happy to write a PR to sample from (some of) these.

[junpenglao]

Sample you mean random generation right? In that case you should pass a pm.Uniform.dist(-2, 2) to init

[kyleabeauchamp]

Yeah, I think we need to add a note about this in the docstring. I spent over an hour tracking this down and assumed I had a botched theano installation...

[twiecki]

@kyleabeauchamp Yes, a note is a good idea.

Note that you can sample (if the GaussianRandomWalk is an observed) by adding nans over the period you want to sample.

[ricardoV94]

Closing this as #4388 seems to fix this issue.

See also #4337 for a general timeseries random issue.