[R-SIG-Finance] Time Series Decomposition

rkevinburton at charter.net rkevinburton at charter.net
Thu Sep 18 23:22:00 CEST 2008


I've been struggling with the following problem and was just 
wondering if someone knew of any solid references in the literature 
for what i'm attempting. I think what i'm attempting is fairly 
straightforward: I have a time series that has a level, a trend and 
seasonal component. I want to model all three of the pieces and put them 
back together to build a forecasting model. I tried 
to use stl but the seasonal component (in the data) is clearly chaging over time so 
the stl models the seasonal piece incorrectly/incompletely (with the amplitude basically fixed for each "season") and this causes all the 
other components to be subsequently modelled incorrectly also. So, I was 
looking for somewhat of a less black boxy approach that 
I can really get my hands around as far as what's going on. Then, with 
that knowledge, maybe I can figure out a way to deal with 
the fact that the seasonal component ( the magnitude ) is changing over 
time. I have heard that I can use DLMs ( dynamic linear 
models for this ) but they are pretty complicated so I was wondering if 
anyone knew of simpler approaches or something in the literature 
that I can read and understand. I have been looking at going the arima 
route but the time series texts that I have don't do in depth coverage 
of a time series decomposition. They talk about how to use the arima 
framework to model a seasonal time series but I have a trend in my 
series also. And I also want to model the level after the seasonal and trend is 
taken out ? So, if anyone could help me out with references for 
what I want to do, it would be greatly appreciated. 
rkevinburton at charter.net

More information about the R-SIG-Finance mailing list