Disclaimer: I am relatively new to time series analysis, so I am not sure if my way of thinking makes sense.

I have a time series of a few years, that I can model relatively well (respectable R squared, and normally distributed errors). However, in the last year of the data set, a series of effects happened that I am almost sure of had a large combined effect on the observed variable. Evidence of this is the fact that for this year, the residuals are no longer normally distributed and show a large upward bias.

I would like to capture this 'combined effect' in the model, by looking at the residuals. I am wondering whether it is possible to regard the residuals as a sum of a signal (of this combined effect) and noise, and to model it accordingly.

Is this a valid method?

  • $\begingroup$ This thing which happened, can you include some variables in your model to account for this? $\endgroup$ Apr 25, 2023 at 15:19
  • $\begingroup$ Welcome to CV, eork. Just as a comment: Perhaps you might study up on 'error correction models' as explicitly modeling part of the data generating process. $\endgroup$
    – Alexis
    Apr 25, 2023 at 15:20
  • $\begingroup$ @user2974951 I have thought about adding a dummy variable, but it's not a constant effect. After it has started at time, say, t, it changes over time. If you have a suggestion as to how to model that, I would love to hear it. $\endgroup$
    – eork
    May 2, 2023 at 12:30


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.