Idgarad
Posts:
20
Registered:
10/18/06


Re: Time Series Analysis Help
Posted:
Dec 7, 2007 9:13 AM


On Dec 6, 2:57 pm, aruzinsky <aruzin...@generalcathexis.com> wrote: > On Dec 6, 1:46 pm, Idgarad <idga...@gmail.com> wrote: > > > > > > > On Dec 3, 7:01 pm, aruzinsky <aruzin...@generalcathexis.com> wrote: > > > > On Dec 3, 3:08 pm, Idgarad <idga...@gmail.com> wrote: > > > > > On Dec 3, 1:21 pm, aruzinsky <aruzin...@generalcathexis.com> wrote: > > > > > > On Nov 20, 5:24 pm, Idgarad <idga...@gmail.com> wrote: > > > > > > > ... > > > > > > coming back to ARIMA(0,1,1)(0,1,1) with a seasonal period of 12 weeks. > > > > > > ... > > > > > > Thirty years ago, I was well versed in time series analysis but have > > > > > forgotten > 90%. > > > > > > What is the significance of the second "(0,1,1)?" > > > > > > According to > > > > > >http://en.wikipedia.org/wiki/Autoregressive_integrated_moving_average > > > > > > , it doesn't belong there. > > > > > The seasonal portion Hide quoted text  > > > > >  Show quoted text  > > > > The notation doesn't ring a bell. I suspect you differenced, 1  > > > B^12, where B is backshift operator, to remove a periodic component. > > > I seem to recall that a common mistake (in my day) is to unnecessarily > > > combine (1  B) to remove a trend with (1  B^n) to remove a periodic > > > compomponent, because any difference, (1  B^k), also removes a trend > > > in the original series. Did you make this mistake? Hide quoted text  > > > >  Show quoted text  > > > (From Duke) > > A seasonal ARIMA model is classified as an ARIMA(p,d,q)x(P,D,Q) model, > > where P=number of seasonal autoregressive (SAR) terms, D=number of > > seasonal differences, Q=number of seasonal moving average (SMA) terms > > () > > > That's what I was referring to, as I mentioned I am on my own on > > learning all this, feel free to slap some sense into me as needed. > > > I have gone back through though and found a calendar of events that I > > can factor into the projection, the seasonality is now 52 weeks rather > > then 12 as there are considerable differences in the activities from > > quarter to quarter. > > > Regardless though I have run into an additional snag (i.e. > > Requirement) is that any model I use I have to backforecast to show > > the accuracy of the model against know existing data. Oh brother, I > > feel like Charlie Brown today. > > > What this is all about is there is a mainframe with different virtual > > computers inside. I have to forecast each virtual computer's usage and > > factor that against capacity to figure out when all hell is going to > > break loose. > > > In short: > > > A : Is production, anything A doesn't use can be borrowed. > > BG : are virtual computers. They get to use a given amount but if > > they need to can borrow A's left overs. > > > I need to learn how to do a seasonally sensitive forecast of AG > > (separately) so I can determine how much they can borrow (if any). Hide quoted text  > > >  Show quoted text  > > The big P and Q doesn't ring a bell. > > I understand that you want an estimator that incorporates seasonally > periodic information, but how far in the future do you want to > forecast? Hide quoted text  > >  Show quoted text 
Me personally, I am only interested in going 2 quarters ahead (roughly 24 weeks) but, if a reasonable amount of accuracy is possible a year at most (Which would allow for some nice "WhatIf" checks). The graph I produce is a sliding 52 week graph so 1/2 of which is existing known data and the second half would be the projections (Thus the last known data point is always in the middle of the graph.)

