GroupVAR: Group Vector Autoregression via Group Lasso

In jeffwong/fastVAR: fastVAR

Description

Fit a VAR model by creating the lagged design matrix and fitting a multivariate response matrix to it. Note that the VAR response matrix omits the first p responses from the input matrix Y. This is because observations in Y are modeled by the p previous values, so the first p observations cannot be modeled.

Usage

  GroupVAR(y, freq = rep(NA, ncol(y)), p = 1,
    weights = NULL, alpha = 0.4, getdiag = T)

Arguments

y	A matrix where each column represents an individual time series
freq	only used if the time series are periodic. freq is a vector of frequencies for each of the time series, as in 'ts(y, freq = ...)'. If the time series are not periodic, then this vector can be a vector of NA
p	the number of lags to include in the design matrix
weights	weights applied to the multiresponse linear regression. Better predictions might come from weighting observations far in the past less so they impact the objective value less. Either NULL, "exponential", or "linear"
alpha	the mixing parameter between group lasso and quadratic, as in 'glmnet'
getdiag	logical. If true, return diagnostics

Details

While multivariate response regressions can be solved as multiple univariate response regressions, this multivariate response problem can better be solved by using Group Lasso. Instead of seeking sparsity in the coefficients for each univariate response, Group Lasso attempts to find sparsity in the coefficient matrix as a whole.

Examples

data(Canada)
  GroupVAR(Canada, p = 3)

jeffwong/fastVAR documentation built on May 19, 2019, 4:02 a.m.