I've studied time series for the past months and I've seen mainly two ways of building a forecasting model:

  • Using ensemble algorithms and making the time series look like a cross-sectional data, in which each column is a lag (t-1, t-2, ..., t-n) and the target is t0. So we would use something like a XGBoost to use the lags to predict the next date. I've actually seen this in a Youtube video from a Kaggle Grandmaster.
  • Using "traditional" econometrics models, like ARIMA. In this case, we would still be using the lags, but in a different manner and first applying some transformation, like differencing, log, etc. This would be done in order to make the time series stationary.

Why we don't take the differences in the first case? Aren't both using the lags to predict the next date value? Why would one make the time series stationary and the other doesn't?

  • $\begingroup$ If you don't need to difference the series to make it stationary, why difference it? If you do, then you should! ARIMA does not require differencing, after all. $\endgroup$
    – jbowman
    Apr 3 at 0:55


Your Answer

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