As many of you know, the google "groups" sci.stat.consult and sci.stat.math were overtaken by spammers. The linkedin.com "groups" has become the place where no spammers can ruin things and great discussions do take place. We invite you to ours. Just google “Autobox” in the groups area to join. Here is a recent post that we put out there to point out some modeling issues with the market leaders that you may find interesting. It really highlights the need and ability to automatically detect and adjust a model for interventions:
“Current practitioners of the seminal work of Box and Jenkins are still getting it wrong!” When Box-Jenkins wrote their landmark textbook in 1967 they analyzed the International Airline Passenger Monthly Time Series(ie known as the "Airline" series and also as the "Airline model") they had a highly seasonal upward trending series. By the way, this is the most studied time series on the planet.
They needed a way to 'detrend' the data to make it 'constant' as that is necessary in order to gauge correlation within the data. They used a LOG transformation. They didn't know how to look for outliers and missed the fact that the last year had a few that skewed their view. If you plot the std dev vs the mean of each year for each of the 12 years "bucketed by year" of data you will CLEARLY see the last year is out of whack. There was an outlier in March(too low) an outlier in July(too high) and an outlier in October(too high). This deviation causes the conclusion to be that the std dev is increasing with the mean which is false if you look at the plot and that you need to transform the data. A common tool is the LOG transformation. Its also nice as the coefficients are the elasticities so economists swarm to this convenient and often incorrect usage.
So, Box-Jenkins didn't know about outliers and when they log transformed the series the forecast was found later to be way too high. In 1973, Chatfield and Prothero( “Box-Jenkins seasonal forecasting: Problems in a case study(with discussion)” J. Roy Statist soc., A, 136, 295-352 attacked the model built by Box-Jenkins as flawed. The forecast had explosive upward growth(you need to transform the forecasts out of logs in order to see this!!!) and was wrong. In 1988, Chang and Tiao wrote a paper “Estimation of time series parameters in the presence of outliers” (Technometric, 30 , No 2, 193-204)that laid out the methodology as to how to solve the problem. We read these papers and adapted to the criticism.
If you go to the SAS documentation or Oracle, xlstat, IMSL, mathworks, Mathematica, Stata and I am sure many others you will see that they all take LOGS. Did they not read the Journal articles? Do they not know the issues?
Here are the links to where you can see others still not truly able to model this dataset even today…
http://support.sas.com/documentation/cdl/en/etsug/60372/HTML/default/etsug_arima_sect056.htm
Note that SAS has a blog on bad practices on business forecasting. Isn't it ironic http://blogs.sas.com/forecasting/index.php
http://oracledmt.blogspot.com/2006/03/time-series-forecasting-2-single-step.html
http://www.xlstat.com/en/support/tutorials/arima.htm
http://www.vni.com/products/imsl/documentation/fort06/stat/NetHelp/default.htm?turl=bctr.htm
http://www.mathworks.com/products/statistics/demos.html?file=/products/demos/shipping/stats/stattsdemo.html
http://media.wolfram.com/documents/TimeSeriesDocumentation.pdf
http://www.stata.com/bookstore/pdf/arima.pdf
Even Teachers at University of Missouri-Kansas City!!!!!
http://forecast.umkc.edu/ftppub/m7stud/mini7-7.doc
You may leave the list at any time by sending the command
SIGNOFF allstat
to [log in to unmask], leaving the subject line blank.
|