Time series analysis tutorials?

Questions
1

I was reading this tutorial from Thomas Wiecki and I really enjoyed it.

Appart from the Stochastic Volatility Model tutorial, do you know of any other relevant tutorials for time series analysis with PyMC3? I can’t find a lot of information on the net on how to use a lot of the distributions in Time series distributions.

Thanks!

2

Thanks! There are no other tutorials but that’s a great suggestion. I put it on my blog-posts-to-write list, not sure when I will get to them though.

Do you have any specific questions we might be able to help with?

1 Like
3

I am still fairly new to machine learning and to time serie analysis so anything would be appreciated, but what I enjoyed about your article was the fact that you transformed a model that didnt realy have any notion of time into a model that was time aware simply by using the Gaussian Random Walk. Not having a lot of experience with time series, I found the concept really interesting. Also comparing the bayesian way of doing things with the frequentist traditional methods was also very interesting.

There was two things that interested me but where I had no idea where to even start. Hiearchical time series where weights vary by time but also per group (where group weights are partially pooled).

Also multi variate time series where multiples things can affect your response variable. For example you are trying to predict store sales over time and you also have time series of foot traffic in front of the store, weather data, etc.

And thank you for all your great articles on your blog! Really useful for a beginner.

4

I was reading this article where he is trying to predict the sales of 811 products having data for 52 weeks for each product. This sounds like a problem where hierarchical modeling could be useful since we have 811 products that should share some similarities and also where we have to account for time for each of them. Also sales from a product could potentially affect sales from another product. This seemed like a really good candidate for bayesian modeling since there’s not that much data per product.