fvarseg

Implements a method for high-dimensional time series segmentation under a piecewise stationary factor-adjusted vector autoregressive model. See

High-dimensional time series segmentation via factor-adjusted vector autoregressive modelling

by Haeran Cho, Hyeyoung Maeng, Idris A. Eckley and Paul Fearnhead. See arXiv:2204.02724 for full details.

Installation

To install fvarseg from GitHub:

devtools::install_github("https://github.com/haeran-cho/fvarseg")

Usage

We can generate an example dataset used in the above paper for simulation studies as

out <- sim.data(n = 2000, p = 100, q = 2, d = 1,
  cp.common = 1:3/4, den.common = .5, type.common = 'ma', 
  cp.idio = c(3, 5)/8, seed = 123)
x <- out$x

Apply fvar.seg with default settings.

fs <- fvar.seg(x, q = NULL, d = 1)

Change points detected from the factor-driven common component.

cs <- fs$common.out
cs$est.cp

Visualise the statistics involved in the multiscale moving window-based procedure for detecting change points in the common component.

par(mar = rep(2, 4), mfrow = c(2, 2))
for(rr in 1:length(cs$G.seq)){
  matplot(cs$est.cp.list[[rr]]$norm.stat, type = 'l', xlab = 'time', ylab = '', main = paste('G = ', cs$est.cp.list[[rr]]$G, sep = '')) # change point detector from each frequency
  lines(cs$est.cp.list[[rr]]$stat, col = 4, lwd = 2) # aggregation
  abline(v = out$cp.common, col = 2, lty = 3) # true change points 
  abline(v = cs$est.cp.list[[rr]]$cp, col = 4, lty = 3) # change point estimators 
  abline(h = cs$est.cp.list[[rr]]$thr, col = 3) # threshold
}

Change points detected from the idiosyncratic VAR process.

is <- fs$idio.out
is$est.cp  

Visualise the statistics involved in the multiscale scanning procedure for detecting change points in the idiosyncratic component.

par(mar = rep(2, 4), mfrow = c(2, 2))
for(rr in 1:length(is$G.seq)){
  plot(is$est.cp.list[[rr]]$stat, type = 'l', xlab = 'time', ylab = '', main = paste('G = ', is$est.cp.list[[rr]]$G, sep = '')) # change point detector 
  abline(v = out$cp.idio, col = 2, lty = 3) # true change points
  abline(v = is$est.cp.list[[rr]]$cp, col = 4, lty = 2) # change point estimators
  abline(h = is$est.cp.list[[rr]]$thr, col = 4) # threshold
}