Thread: Time series Help

Hiya ,

If anyone could help me with a time series prob. Attached is a graph of datas.

Could anyone confirm to me if a multiplicative decomposition model wud be appropriate please.

Could you also help me choosing between a simple or a double exponential smoothing method plz.Explanation wud be very much appreciated.ty very muchx...

Thank you guysx

Cheers,
Amazed

Originally Posted by Amazed

Hiya ,

If anyone could help me with a time series prob. Attached is a graph of datas.

Could anyone confirm to me if a multiplicative decomposition model wud be appropriate please.

Could you also help me choosing between a simple or a double exponential smoothing method plz.Explanation wud be very much appreciated.ty very muchx...

Thank you guysx

Cheers,
Amazed

I'm not sure that there is significant seasonal or periodic components in
this time series. In which case multiplicative or additive decomposition are
not relevant (are they?).

As you have a trend in the data, double exponential smoothing is appropriate.

RonL

Choice of gains in double exponential smoothing:

This reference suggests these are selected using non-linear least squares using Marquart
(presumably their shorthand for the Levenberg-Marquart algorithm)

My take on this is that you choose them by trial and error.

But given what the gains are used for I would go for (1-gamma) ~ 0.3, and
alpha ~0.2 as a first guess, if that proves unsatisfactory, mabe reduce (1-gamma)

(gamms and alpha are as in the notation of the reference)

RonL

Originally Posted by CaptainBlack

Choice of gains in double exponential smoothing:

This reference suggests these are selected using non-linear least squares using Marquart
(presumably their shorthand for the Levenberg-Marquart algorithm)

My take on this is that you choose them by trial and error.

But given what the gains are used for I would go for (1-gamma) ~ 0.3, and
alpha ~0.2 as a first guess, if that proves unsatisfactory, mabe reduce (1-gamma)

(gamms and alpha are as in the notation of the reference)

RonL

Some trial and error suggests that the initialisation of b1 is more important
than the gains in getting a good result here. In the attached SS I initialse
b1 to the avaerage slope in the data and get reasonable results for quite
wide ranges of alpha and gamma. Try it and see.

RonL