Assume a time series with a clear seasonality with observations every quarter. If you want to use that time series and make predictions four steps ahead, but you are only interested in what the forecast say on an annual level (mean of the four quarters for example). Is it in this case important to remove seasonality before modeling?
$\begingroup$
$\endgroup$
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
2
In forecasting, one typically does not remove seasonality, but models it with a seasonal model. The tag wiki contains pointers to literature; I especially recommend the online textbooks by Athanasopoulos & Hyndman.
In a situation like yours, there is no hard-and-fast rule about what method will yield the best forecasts.
- You could forecast with a seasonal model (e.g., Exponential Smoothing or ARIMA) and then aggregate.
- You could also decompose your series (e.g., using STL) and only forecast the trend and the level component.
- You could aggregate your series to years and then forecast that series with a non-seasonal model.
- Or, finally, you could do all three methods and then reconcile the forecasts, e.g., using MAPA.
Which approach performs best will depend on your time series, although the last approach using reconciliation typically outperforms single methods - but it's most complex, too, of course.
answered Sep 20, 2022 at 7:26
-
$\begingroup$ @Stephen Kolassa Thanks. Does it makes any difference if I opt for a linear regression model in contrast to a classic univariate time series model like ARIMA, or is the four points you highlighted above valid for both approaches? $\endgroup$– HenriCommented Sep 20, 2022 at 7:39
-
1$\begingroup$ You mean a regression on the quarter? That can indeed be competitive, depending (again) on your data. My points would also apply to that. You could even add such a model to the three first bullet points above, and reconcile all four separate forecasts. $\endgroup$ Commented Sep 20, 2022 at 9:07