Popularity hide and seek | PollsPosition — The Blog

[utterances-bot]

Popularity hide and seek | PollsPosition — The Blog

Estimate latent presidential popularity across time with a Markov chain

https://alexandorra.github.io/pollsposition_blog/popularity/macron/hidden%20markov%20models/polls/2021/05/16/hmm-popularity.html

[jroberayalas]

Hey Alex! This is one of the best posts I've encountered about gradually implementing a Bayesian hierarchical model. I really had a lot of fun going through it. Kudos! A couple of questions I had while reading through it:

• Suppose the presidential duration goes beyond 60 months, how can you continue the forecasts for consecutive months beyond 60? Once the model has been trained, is there a way to keep pushing the Gaussian walk into the following months?
• And related to my previous question, will this forecast continue the trend from the last months, or does a new term need to be added to the model to capture the trend, e.g. b0 * time?

Great work again! Looking forward to keeping the discussion.

[jroberayalas]

Hi again! Just realized that the issue of forecasting beyond month 60 not only affects the Gaussian walk but also the month effect component. Given that we have a Gaussian distribution for each of the 60 months, the moment we go beyond that time frame, we are left without information and therefore distributions to model. Hope this makes sense. Not sure if there's an approach that can propagate the month-related components of the model into the future.