| Nonstationary Time Series Analysis and Decomposition using Time-Varying Parameter Autoregressions |
This site provides software for fitting, analysis and exploration of time
series using classes of time-varying autoregressions -- or TVAR models. In addition to model
specification, selection and assessment, the software develops time series decompositions to
explore underlying latent component structure in observed data -- a general and flexible time-domain
approach to "time:frequency" decompositions to elucidate patterns of change over time in frequency structure
of nonstationary series.
Feel free to download and explore, and let us have your feedback as
we update the material. Any identified bugs will be added to our
bug sheet so please check there first before contacting us.
Downloads:
- Fortran90 and S-Plus software
- Documentation
This includes basic details of software, inputs, outputs and post-processing, with examples.
The document describes the use of the Fortran90 and S-Plus routines (the current
version, updated recently to correct a few typos, is dated July 27, 2000).
- Matlab software implementing precisely the same models and methods. The Matlab
routines are preliminary versions and will be updated periodically. Though the above
documentation concerns the Fortran and S-Plus software, the use of the Matlab functions
is very similar. Examples are given in the Matlab directory (see below).
The software is bundled here
as a zipped unix directory (the current version is dated July 27, 2000). Download and create the
directory tvar using the unix command sequence
gzip -d -c tvar.tar.gz | tar xf -
This will create two subdirectories: tvar_f90_splus and tvar_matlab.
Alternatively, you can directly download the the full directory structure and individual
files (unzipped) from here
Using Fortran90 and S-Plus software:
When downloaded, you will find that the directory tvar_f90_splus includes several Fortran90 files, an S-plus
file, some additional files with data etc for the examples in the documentation,
and the important Makefile. The Fortran90 code uses the public domain LAPACK software
for linear algebra, so this must be available on your system.
There are three main Fortran programs: grid90.f, tvar90.f, decomp90.f To compile these simply use
the unix commands
- make grid90
- make tvar90
- make decomp90
Using Matlab software:
The Matlab routines are free-standing. In the tvar_matlab directory is a short README file
and the file tvar_eg with some examples of how to use
the functions. You may freely edit the .m files to alter and
customise these functions for your own use. We do plan to develop them
further and updates will be posted on this web site as they arise.
Key references:
-
Latent Structure in Non-Stationary Time Series.
(1998) R. Prado, PhD thesis, ISDS, Duke University
-
Exploratory modelling of multiple non-stationary time series:
Latent process structure and decompositions.
(1997) R. Prado and M. West,
In Modelling Longitudinal and Spatially Correlated Data,
(ed: T. Gregoire), Springer-Verlag.
-
New methods of time series analysis for
non-stationary EEG data: Eigenstructure decompositions of time varying
autoregressions.
(1998) A. Krystal, R. Prado and M. West,
Clinical Neurophysiology, 110.
-
Evaluation and comparison of EEG traces:
Latent structure in non-stationary time series.
(1999) M. West, R. Prado and A. Krystal,
Journal of the American Statistical Association, 94.
-
Time series decomposition.
(1997) M. West, Biometrika, 84, 489-494.
-
Bayesian Forecasting and Dynamic Models
(1997 - 2nd Edition), M. West and P.J. Harrison, Springer-Verlag, New York.
Key material is covered in Chapters 9 and 15.
- Smoothness Priors Analysis of Time Series.
(1996) G. Kitagawa and W. Gersch.
Lecture Notes in Statistics #116, New York:Springer-Verlag.
Key material is covered in Chapters 11 and 12.
Enquiries or question to raquel@ams.ucsc.edu or
mw@stat.duke.edu
Research underlying the software presented here was performed under partial
support from NSF grants DMS-9704432 and 9707914
This software is made freely available to any interested user. The authors can provide
no support nor assistance with implementations beyond the details and examples here, nor
extensions of the code for other purposes. The download has been tested to confirm all
details are operational as described here.
It is understood by the user that neither the authors nor Duke University bear any responsibility
nor assume any liability for any end-use of this software. It is expected that appropriate
credit/acknowledgement be given should the software be included as an element in other software development
or in publications.
|