I have a basic knowledge on econometric techniques for forecasting, and I was reading about Forecasting with decomposition.

My data seems to show high seasonality as can be seen in the following picture:

enter image description here

If my understanding is correct, after decomposing the time series in its components with (e.g.) STL, I can forecast the seasonally adjusted component and then sum the season_year to the results to obtain predictions in the original scale. Is this correct?

I am using Machine Learning (Random Forest precisely) to perform the forecasting, and my questions are:

  1. Does decomposing the time series like I am doing make sense when using ML?
  2. Since I am using other covariates to forecast, do I need to decompose them as well and use their seasonally adjusted values instead of the original ones?

In the past I used to add dummy variables like quarter(date) to my model to account for the seasonal effect, but maybe this is better.

  • 4
    $\begingroup$ Your data doesn't look seasonal at all. What makes you think it is? $\endgroup$ Jul 24, 2023 at 12:06
  • 2
    $\begingroup$ Note that the "seasonal" component of an STL decomposition will fit a seasonally-looking component regardless of the data, and that in your case this component is on a scale that is about 10 times smaller than the original data. As Rob writes, your original series is absolutely nonseasonal. Don't just look at the "seasonal" STL component! $\endgroup$ Jul 24, 2023 at 12:09
  • $\begingroup$ Thanks for these info. Indeed, I was probably misled by the 3rd plot without taking into account its scale. $\endgroup$
    – umbe1987
    Jul 24, 2023 at 12:16


Your Answer

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