A Data Science Central Community
This came in my mailbox as a sales pitch by Autobox, however I thought that it is interesting:
Since we are always interested in learning about how others do time series and testing how our approaches work vis-à-vis other dated procedures, we pursued the data and would like to share our results.
Sometimes in an educational process, simplifying and simply wrong assumptions are made to solve complex time series. Unfortunately, students take away the idea that methods and approaches that were correct for a textbook example can be carried over to any data set when in truth they may be doing damage.
We saw a Blog discussing "The Dynamic Stability of AR Models - Tricking EViews".http://davegiles.blogspot.com/2012/01/dynamic-stability-of-ar-model...
This post isn't really about Eviews, but focused on the standard and outdated approaches that fail to truly model the data. Assuming a log transformation to deal with non-constant variance is at the heart of why we disagree with those that do it "this way". A lot of built-in assumptions or more precisely built-in transformations are placed upon the data yielding inferior results.
We downloaded the example data and ran it in Autobox automatically and posted the results in the comments section with an explanation of the differences(look for my name in the comment section). The bottom line is that after applying the Chow test and identifying a break in the parameters and deleting the first half of the dataset there is no proven change in variance and no need to assume a log transform.
After reading the blogspot post, check out another log example that we found. Again, the series is NOT increasing variance in the last years as it has stabilized. It too can be remedied using the Chow test to truncate earlier data.http://www.stat.pitt.edu/stoffer/tsa2/R_time_series_quick_fix.htm