Advice for Time Series Forcasting

Questions version agnostic
5

DFA would be quite easy to write as a statespace model, PRs welcome to pymc-extras to add that. The model is covered in Durbin and Koopsman Chapter 3.7

In general I agree with everything @JAB said above. I’ll just add 2 things:

  1. If you have plateaus, decreases, and short-term improvements, it might imply an HMM where there is a discrete latent state governing the time series dynamics
  2. The statespace module unfortunately don’t scale well at the moment, so if you have thousands of time series that you want to jointly estimate, it’s not a good choice. I hope this will change in the near future.
  3. For cases with big data and many time series, I suggest a method that reduces to linear regression: prophet, splines, or HSGP. Recursive models are always going to be way more computationally intensive than y = X \beta, even if getting X requires some effort.
2 Likes