Time series, wallowing... After ARIMA then,,,

Skip to first unread message


Nov 22, 2013, 1:04:38 PM11/22/13
to pystat...@googlegroups.com
As is evident from my last post on ARIMA I'm searching for various approaches to Time series forecasting,, It seems that ARIMA won't do what I'm looking for.
I have a pretty good method that is yielding good results (maybe by chance but the results are pretty good).

Essentially I take the percentage difference of the sales in the 4th quarter from the prior year and apply that to the most recent full year and then divide the full year projection that results bu the average percentage of sales that each month has gotten historically. Across several markets my worst to date error is about 6% with most markets in the 1-3% error range.. Thi is all done with manual spreadsheets (Somewhat painfully) here (public) https://docs.google.com/spreadsheet/ccc?key=0Ak1ecr7i0wotdEJfcXhiTWZHSUZlWjMzUnM1Mk1WTWc&usp=sharing Note many tabs.. This approach takes off from this post http://nomascusconcolor.hubpages.com/hub/Bayesian-Forecasting-Methods-for-Short-Time-Series which claims to be Bayesian but Idon't see that!

This is also the data that I asked about in my prior question on getting ARIMA to work.. I now get that the better and currently unavailable method should be SARIMA, but I suspect that even that would not be as accurate. I'm not desperate to use these methods but rather am using the fact that I understand this data domain pretty well and therefore it is useful to try using different methods to learn them..

In reading I came across this slideshow from SCIpy2011 (McKinney, Perktold, Seabold) http://www.slideshare.net/wesm/scipy-2011-time-series-analysis-in-python And slide 26 has a preview DLM model that look very much like something that could fit my data. I can't find any subsequent reference to anything related?

Anyhow I'm trying to write up what I've learned and where I'm still confused (more the latter) Which will at least serve to link resources I have located on the web.

This is the notebook Chad Fulton redid so nicely on my former ARIMA question, modified with Data linkages and all imports http://nbviewer.ipython.org/7604008  As one can see the forecast in the last cell does not reflect the seasonality nor the upward drift from the ecconomic cycle. Having fiddled with the ARIMA params for some time I don't seem to be able to correct that.

So Ive rambled a bit, but the question is if any of you have advice or guidance, ideas or code, I'd like to believe that carrying this out more could be a useful example for highly seasonal time series in general and hopefully the python tools to carry it out with? The papers and docs are usually so dnes with the math it is hard for the non-stats majors such as myself to apply, so well worked examples are very valuable (and rare!)
Reply all
Reply to author
0 new messages