.random() methods missing from some timeseries distributions

[michaelosthege]

Timeseries distributions with a .random() method:

• GaussianRandomWalk

Timeseries distributions where .random is None:

• MvGaussianRandomWalk
• MvStudentTRandomWalk
• AR
• AR1
• GARCH11
• EulerMaruyama

At least for the two Mv distributions is should be straightforward to implement.

Make sure to include them in test_distributions_random.py, just like GuassianRandomWalk already is!

[awray3]

Hi, I am new to contributing and would like to work on this issue. Can I work on it?

[michaelosthege]

@awray3 of course, go ahead!
Just make sure to mention this issue when opening a PR, so it gets linked.
Also check out the wiki pages. We have some code style guidance there.

[awray3]

Is MvGaussianRandomWalk intended to only take one of cov, tau, or chol, much like how GaussianRandomWalk only accepts one of tau and sigma but not both?

Also, are each of the random walks conventionally taken to start at 0 like GaussianRandomWalk?

[AlexAndorra]

Is MvGaussianRandomWalk intended to only take one of cov, tau, or chol, much like how GaussianRandomWalk only accepts one of tau and sigma but not both?

Yes, these three parameters are just three different ways of expressing the covariance matrix. You can take a look at MvNormal to get a better idea if needed.

Also, are each of the random walks conventionally taken to start at 0 like GaussianRandomWalk?

Yes, I think so. I'm guessing it's like Gaussian Processes: you can just add the mean separately afterwards to center the random walk on something else than 0

[awray3]

@AlexAndorra Thank you for the pointer; it lead me to _QuadFormBase which had exactly what I needed.

[awray3]

General question on time series in PyMC3, which I could not find in the documentation: what should the shape parameter be for a multivariate time series random variable? My guess after sorting through some of the code was that it should be something like shape = (time, n_dims), but I found no definite answer. I will consider adding shape to the docstrings for these functions to eliminate any guesswork.

[mjhajharia]

Entire time series is being refactored so maybe we can close this @michaelosthege ?