Re: Help needed for forecasting model

#1
Hi,

I have a dataset of some monthly usage values. I have to forecast the for the next 12 months. I have taken log of data for ease of use. It seems to me that ARMA(2,0,2) model fits best for it. Can you confirm if I am right or wrong? If it is, please tell me which one fits the best.
Also, I have used JMP and NumXL for doing the analysis and both give me different results. NumXL supposedly does some optimization of parameters. But I am not sure which one to believe. Can anyone tell me how to optimize different parameters?

Data is attached.
 

chiro

New Member
#2
Hey hrushiayre.

The key thing for your situation is to formulate a hypothesis and then with the appropriate distribution and test-statistic, you test it.

Doing this requires not only statistical knowledge but also domain knowledge.

The simplest hypothesis test would be testing one parameter is equal to something with the alternative being something is not equal to that same value.

If you have three independent parameters (given the ARMA(a,b,c)) then one suggestion I would have is H0: a = 2 and b= 0 and c = 2 vs H0: one of those values is different (this is assuming that using an ARMA model is sufficient which I don't know if it is or not).

In terms of assessing different kinds of fits you should look at specific fit statistics (like AIC and log-likelihood) and in the right context, make comparisons between these fit statistics in the context of your problem.