1

I have a time serie, and I want a stationary process for search posible models. One of the requirments is normality.

shapiro.test(serie)
p-value = 0.0002322

How can I normalize my time serie? (I try with Logaritimic transformation, regular differences, stationary differences, but doesn't work)

It's necesary normalize the serie?

When I look the p values for Ljung Box they are not normals. This is because the not notmality of the serie?

shapiro.test(model1$residuals)
p-value < 2.2e-16

EDIT: model$residuals and p values enter image description here

user4563174
  • 25
  • 7
  • 1
    Is it the residuals from a model or the actual series that you're attempting to model (dependent variable or output variable, etc.) that you want to normalize? Are you sure that you require the latter and not just the former? – Graeme Walsh Jan 16 '16 at 19:14
  • @GraemeWalsh I edit the post, I wish that it was the information that you ask – user4563174 Jan 16 '16 at 19:18
  • 1
    Regarding your first paragraph, normality is not a requirement for stationarity. Why do you wish your series to be normal? – Richard Hardy Jan 16 '16 at 19:52
  • 1
    OK, so it's the residuals that you want to be normal. I presume this requirement stems from that fact that you're assuming the errors are random variables and distributed normally (say, for purposes of ML estimation). If this is the case, you may just need to handle the (four) large residuals by means of appropriate outlier and break detection. There could be other reasons, however. Transforming the residuals directly is not the way to achieve normality - you need to make changes to your model then see how the estimated residuals are affected. – Graeme Walsh Jan 16 '16 at 20:02
  • Very good point by @GrameWalsh: transforming the residuals is not the way to go. Your goal should be to achieve residuals that have desirable properties as they are, not after any transformations. And for that you would change your model, just as he indicates. – Richard Hardy Jan 16 '16 at 20:07
  • @GraemeWalsh I concur with your remarks and suggest that http://stats.stackexchange.com/questions/190436/box-cox-transformation-makes-out-of-sample-forecast-error-worse be looked at as it is I believe relevant to this discussion – IrishStat Jan 16 '16 at 21:39

0 Answers0