I figured out I should make my point as of furture
improvements of JMulti. It is already a great tool, but
to the extent something need to be improved, I'd like
to add my take on what should be the focus of new
Broadly speaking, this is just one phrase: INSIST ON
STABILITY ANALYSIS !
The widespread practice of econometrics consists of
estimating a model and then draw inference on the basis
of the results, most of the time with reference to
economic theory. Essentially two options are possible :
1- The inference is derived from a SINGLE specific
coefficient. This is the case of unit roots for example.
2- The inference is derived from a set of interelated
coefficients. This is the case of Granger causality or
IRFs for example.
My metaphysical problem with this is that the inference
drawn from cases 1 or 2 is very limited. How can you
not tell that the process you are studying is not
stationary during a certain time, and then becomes
non-stationary ? In any case estimation of breaks in
the process, as widely known, is an important
prerequisite. The same applies to the multivariate case
: how can you tell the result of granger causality you
have just found on the whole sample is true all along
the sample ?
Sure, some break tests are already available in JMulti.
But I should encourage a better visibility. I also wish
an implementation of break tests in the multivariate
context. I should stress that point because JMulti is
oriented towards the multivariate case.
Besides break dates, I also encourage the developement
of automated rolling estimation procedures. And by
rolling I do not merely mean recursive estimation, but
sliding sample. I don't know if there is a theory yet
for that, or even if it is needed. I believe it
important to check whether the results you get on a
specific sample still hold on another sample, or parts
of the sample. This is what economists are looking for.
Also inference in the multivariate context should
emphasize robustness checks, because practical people
want a robust result. The mutlivariate case, like VECs,
is also subject to wide result variance, because the
number of parameters to estimate is high. Change the
lag length for instance and your results are likely to
change. Now I don't know what you can do about that ;
maybe a dedicated help section on the importance of
parameter choices would help a lot.
I know some steps are already being followed in some of
the directions I mentioned. I know you can manually do
the things I have stressed. Yet I also know that
economists are not econometricians (like me), and if
only for that reason, a whole built-in stability
analysis is required.
I hope this quick note will help build a better JMulti
environment with economists's satisfaction as a goal.
In the meantime JMulti 4 is great, so keep going !
Log in to post a comment.