I have recently been tasked in developing some forecasts for some CPU usage.
I currently have found data from 2003 to current. The data is weekly on a Julian calendar (52 weeks a year). I have also found a calendar of holidays and planned deadlines and activities that may affect the usage (they are dummy variables I suppose, with a 1 indicating yes and 0 for no.) Now after about a month of pouring through every scrap of online lore I could find, and with some assistance from several people I have a slightly better idea of what to do. I have to do a time series analysis of the CPU usage in relation to seasonal activities (seasonality being 52) along with the the dummy variables indicating various activities that affect the cpu usage. As I am not (big not here) a statistics guru (I'm a a techie not a mathematician) and this is more for the sake of research to make my life easier. I have $0.00 to dedicate to this (I can't even get a book out of this...) so Open Source is my only option. I would like to use Gretl to develop a 52 week forecast based on past historical data. I have the date (actual and week #) and the known cpu usage in MIPS (million of instructions per second). So far: Date,MIPS Then I have anywhere from 10 to 300 (depending on what I want to add) columns of potential events that could affect the MIPS. (dummy variable, 0 or 1, holidays, weather, hell even the price of oil if we choose). After all this research ARIMA (Box-Jenkins) seems to be the system to use. I have used Autobox to get familiar with this but there is no way to back-forecast in that tool that I can see (We want to be able to check the model against older data for accuracy and graph if) and I would really like to have a better understanding of this process. Ideally I would like to have this as automated as possible (that's what I do for a living) but not having the strong backgound in statistics I am trying to automate something I barely understand. Each week I get a new data point to add to my master list so we want to develop a script\pipeline\process and need to re-run this weekly (plus perhaps make some adjustments in future calendar entries based on changes). (in short back-forecast 52 week including now and give me the next 52 weeks, trim graph to that period...) But, lost, alone, and frightened in the dark woods of statistics I need help building this process using Gretl (and if necessary R). But where do I begin? How should I structure such a script for Gretl? Ok I have Gretl... Check! I have R also... Check! (just in case) I have all the data ready to go in excel files (that's how I get them)... Check! How do I put this puzzle together? How, in Gretl do I determine the best ARIMA model through a script? Do I just loop through? Suggestions? R appears to have some functionality for autofitting, what process should I go through via Gretl to find things like this out script wise. I'm lost and I hear wolves and strange noises coming from the forests (sounds like albino calculus monsters... I heard the differential ones are poisonous...) I know there has to be others out there that are in the same boat and I am hoping to put an easy guide together so others can use it as a template. Where do I begin? How do I know when to end? How can I use Gretl to accomplish this goal? Help! The fire is dying and I hear rustling in the bushes nearby! (and for the Jetson fans HEPL!) Id.
