jeffwong/fastvar

R package for Vector Autoregression

fastvar's Issues

multivariate lm is not parallelized

The native lm call on a response matrix just makes multiple calls to lm on a univariate response vector. VAR models can be computed more efficiently by solving the normal equations in parallel

parallel SparseVAR should use parallel instead of multicore

R >= 2.14 has a built in parallel library which inherits from snow and multicore. This will allow parallel computation of glmnet on both windows and linux

consider group lasso in SparseVAR

http://www-stat.stanford.edu/~tibs/ftp/sparse-grlasso.pdf

glmnet cannot suppress intercept term

This means SparseVAR and SparseVARX will not be able to take intercept=F

predict with seasonality yields NA

first few values are correct but next few values are NA

use glmnet mgaussian family

allows for multi response lasso regression

combine VAR model with HMMs

implement time varying coefficients

implement L2 penalty for wide design matrices

implement efficient updates to VAR coefficients when changing weights

The weighted least squares problem has the normal equations

(X^T W X)^-1 X^T W Y

implement graphical lasso for clustering of time series

reimplement VAR and VARX using GSL

use GNU scientific library to implement faster versions of VAR and VARX

use structures to move parameters around

incorporate k-means preclustering into y.spec

Create a helper function to autogenerate the specification matrix for SparseVAR and SparseVARX. These matrices are used to determine which endogeneous variables are regressed on which other endogenous variables. This can be determined by clustering similar time series together

Estimating MA params

http://www2.econ.iastate.edu/classes/econ674/bunzel/documents/Lecture4.pdf

solve var using weighted OLS

notes

http://faculty.washington.edu/ezivot/econ584/584notes.htm

wrap fastVAR data structures around zoo

Have built in methods for decomposing time series

Preprocess data by removing seasonality and trend. See Rob Hyndman's decompose function in package forecast

time series intervention

http://www.r-bloggers.com/time-series-intervention-analysis-wih-r-and-sas/