0

I am trying AUTO ARIMA model on a dataset which has no trend and no seasonality. When I run it, it will forecast a flat line from p,d,q value selected from AIC. But when I take another approach, where I iterate for a range of p,d,q and pass it to ARIMA, it shows better result with variation in forecast. It may have high AIC, but it was clearly a better forecast. I cant do this manual ARIMA because I have to forecast 1000+ timeseries (small datasets with 100 data point). Is there any way we can extract p,d,q in AUTO ARIMA based on high variance in forecast result?

Ex:

Below Image is result of AUTO ARIMA. Green line is forecast

enter image description here

Below image is result of Manual ARIMA. Orange line is forecast (here test data is not shown)

enter image description here

Shravan K
  • 21
  • 3
  • No. It sounds like your goal is to overfit as much as possible rather than produce good forecasts, so just set $p$, $d$ and $q$ to arbitrarily high values and it should work pretty well. – Chris Haug Jun 23 '21 at 16:09
  • i want good forecast. But i want to show some variation in forecast instead of a flat line. When u say high value, u suggest pdq>3,4,5? – Shravan K Jun 23 '21 at 16:32
  • Why do you think a variable forecast is better than a flat one? [Very often, a flat line is the best forecast possible.](https://stats.stackexchange.com/q/124955/1352) – Stephan Kolassa Jun 23 '21 at 16:51
  • This question could benefit from more detail about how you evaluate a forecast to be "good" (it doesn't appear to be the same definition that most people use). For example, if you could include data and the result of `forecast::auto.arima` as well as what you consider a "better" forecast for the same series. – Chris Haug Jun 23 '21 at 22:06
  • I have updated the question Chris – Shravan K Jun 24 '21 at 03:30

0 Answers0