Contents
|
Cointegration & Common Trends
Cointegration & Common Trends Now how can I explain the idea of cointegration without getting too technical... (By the way, do you like the eye glasses; I put 'em on to make me look schmarter!) Well, imagine the case when one variable (your yearly income) depends on two factors, your day job and your night job. The first factor (your day job income) increases by, say, 5 percent per year on average, while your night job pays you roughly the same every year. Suppose that you save 10 percent of your yearly income from your day job, and spend all the money you get from your night job (at your local pub). Then both your total yearly income and your yearly savings are trending (or "integrated" given appropriate mathematical assumptions), while your yearly savings minus 10 percent of your yearly income (equal to minus 10 percent of your night job income) is fairly constant from year to year. Accordingly, your yearly income and savings trend for the same reason (they have one common trend), while a particular combination of the two remains quite constant (or is cointegrated.)
Remove me!
Markov Switching VAR's
Markov Switching VAR's Still wearing them spectacular eye glasses! Whaddaya think? Pretty cool, huh? Should I get a pair like that... When I first tried to visualize a Markov switching VAR model, I had the picture of a volcano in mind. Most of the time it's kind of quiet, occasionally letting off some steam, but with little variability in terms of seismic activity. (In case you're wondering, I know nothing about geology.) Then, every once in awhile, it's behavior changes dramatically (sort of like the 3rd movement of Alan Hovhaness' "Mount St. Helens" symphony - check out the recording on Delos with Seattle Symphony and Gerard Schwarz). Now, suppose you want to predict the behavior (seismic activity) of a fairly active volcano. It's probably true that a linear VAR (or AR) model will do a pretty good job except when you really want it to. Once the volcano goes into a pre-eruptive state, a different (linear) model may be better able to predict its behavior. In other words, the (optimal) weighting of the past and current information is likely to have changed once the volcano is in such a state relative to when it's in a state of "sleep". For instance, past pre-eruptive (and eruptive) times will probably be given higher weights, while the periods of sleep may be completely uninformative.
Remove me!
GAUSS & RATS CodeYou can download zip files with my code for estimating Common Trends and Markov Switching VAR models below. The common trends code was originally written for RATS version 3.x and it has been updated for RATS version 4.x by my good friend and colleague, Henrik Hansen, at the Department of Economics and Natural Resources, the Royal Veterinary and Agricultural University in Copenhagen. Both versions are available for download. The code for MSVARs, on the other hand, is written for GAUSS version 3.2.x and can be run on both the MS-DOS (16-bit) and Windows 9x/NT (32-bit) versions of GAUSS; it will probably also work for all operating systems that support GAUSS. If you find any bugs, please report them to me so that the code can be corrected. Downloads
Readme files are included in all packages and I advice you to read them before you attempt to run the code.
[Home]
Last Updated: August 29, 2000 |
|||||||
|