1

May data is structered in the following way:

Week    Sales
10
9         
8         
7         6
6         8
5         5
4         4
3         5
2         7
1         6

Therefore, when I want to make a forecast for week 10, I don't have the information of the number of sales in week 8 and 9 due to logistics constraints.

Is there a way to adjust an ARIMA model when you have a gap in your data like this?

What should I do? Thanks

Maria
  • 21
  • Do you have reason to believe an ARIMA model would do well with your problem? ARIMA's have a pretty bad track record at prediction compared to much simpler methods, especially for the near term predictions. Winters or Holt's method does as well or much better in many instances, as does exponential smoothing. You typically need LOTS of data to fit an ARMIA model, and then hope that the autocorrelation structure remains intact into the future as well. –  Jan 14 '14 at 00:04
  • @Eupraxis1981 The problem is that I am working with weekly data, with a periodicity of 52 weeks, and exponential smoothing models don't work well with such long seasonal periods. – Maria Jan 15 '14 at 09:19
  • With such a long seasonality, I am surprised it would ahve such a strong short-term effect (2 weeks out of 52). Have you tried to de-seasonalize first, then apply a simpler method like Holts, Winters, or expon-smoothing? Try de-trending and deseasonalizing first, then try simpler methods. A 52-term ARIMA will likely not be very robust. –  Jan 15 '14 at 14:10

1 Answers1

0

Why do you need to adjust anything? Just forecast 3 steps ahead.

Rob Hyndman
  • 219
  • Please make this answer more detailed – Lost1 Jan 13 '14 at 08:47
  • But doesn't a model work best if it is adjusted to the right days it is forecasting to? For example, if I have an arima(1,0,1), the model is forecasting taking in consideration the previous observation, which may be poorly related to what happened 3 steps before. – Maria Jan 15 '14 at 09:27
  • Of course the forecasts will be better if you have the additional observations. But if you don't have them, the best thing you can do is forecast 3 steps ahead. – Rob Hyndman Jan 15 '14 at 22:43