Vector Autoregressions?

[benmbrennan]

I’ve been trying all week to figure out how to estimate a VAR using PyMC3. Is there a guide out there for doing this? Most of the priors I’m supposed to use involve a Wishart distribution and I couldn’t really figure out how to map that into the LKJCholeskycov distribution, nor could I figure out how to take that and use it in an MV Normal distribution.

[junpenglao]

Are you using the multivariate timeseries like pm.MvGaussianRandomWalk, maybe it is easier if you can show your code.

[benmbrennan]

Honestly at this point it’s all nonsense from trying to get things to work. I didn’t attempt pm.MvGaussianRandomWalk because all of my variables were stationary so I didn’t think that would be appropriate. Maybe I’m misunderstanding the purpose of MvGaussianRandomWalk, but I can’t seem to find any examples of the context it would be used in.

[DanWeitzenfeld]

@benmbrennan - did you end up figuring this out? if so, could you share a small example?

[rhouck]

I’ve posted a new related topic here:

How to implement vector autoregression with interdependencies?

Hello, I’m hoping someone can help me implement a vector autoregression in pymc3 that includes interdependencies between time series. vector autoregression described here: https://en.wikipedia.org/wiki/Vector_autoregression I understand I can use pm.AR to model multiple time series simultaneously, however it seems the rho parameter is limited to two dimensions, implying that time series will be modeled as independent series. I’d like to set up a model that estimates weights on a series’s own …

vector autoregressions of independent series are doable by passing a 2d set of priors for the rho argument to pm.AR. I’m not sure how we let these series load on each other’s lagged values though.