detectChangePointFast: Detect Change Point Faster

In speegled/cpbaywave: Bayesian Wavelet based Change Point Detection

Detect Change Point Faster

Description

Applies Bayesian-wavelet technique to determine whether a multi-dimensional time series has change point in the mean. Returns the 3-5 most likely change point locations.

Usage

detectChangePointFast(
  a,
  setdetail,
  useBFIC = TRUE,
  showplot = FALSE,
  showall = FALSE,
  padding = "insertion",
  slow = FALSE
)

Arguments

a	A vector or matrix representing time series. If matrix, each row is the value at a single time. Numerically unstable if data dimension is greater than about 50. Use JLDetectChangePoint in that case.
setdetail	The detail levels of the wavelet transform to use to detect change in mean. Default is all levels.
useBFIC	Set to true to choose the change point with highest BFIC.
showplot	Set to true to see a plot of the probabilities of a change point at each time, together with a scatterplot of the first dimension versus time.
padding	One of mirror, extension or insertion. Default is insertion.
slow	Set to TRUE if you want to check each point separately, rather than fast search for change point.

Value

value The value of the BFIC or maximium probability if useBFIC = FALSE. BFIC greater than 3 is evidence that there is a change in mean.

index A vector giving the 3-5 most likely (or highest IC if useBFIC is TRUE) indices where a change point occurred.

Examples

a <- createTimeSeries() #True change point at time 72
detectChangePoint(a, 0:6, showplot = TRUE)

a <- createTimeSeries(mu1 = 1, mu2 = 0, sigma = 1, n = 100, tau = 55)
plot(a[,1])
detectChangePoint(a) #Hard problem. True change point at t = 55. Not always possible to detect.

#Time series with no change in mean.
dim=5
sig1=diag(dim)
mu1=rep(0,dim)
mu2=rep(0,dim)
n=128
tau=70
series=createTimeSeries(mu1, mu2, sig1, n, tau)
plot(series[,1])
detectChangePoint(series)
#value less than 3 indicates no change model is favored


#Time series with smooth mean function with jump.
dim=10
sig1=diag(dim)
mu1=rep(0,dim)
mu2=rep(1,dim)
n=90  #Algorithm doesn't work as well when n not power of 2.
tau=20
series1=createTimeSeries(mu1, mu2, sig1, n, tau)
sn=sin(1*pi*1:n/n) #Compare with sn = sin(3*pi*1:n/n), which is too hard for algorithm
series2=sn+series1
detectChangePoint(series2, useBFIC = TRUE, showplot = TRUE) #True change point at t = 20.


#Time series with smooth mean function, no jump. Illustrates that BFIC unreliable
#when there is a smooth underlying mean function.
dim=3
sig1=diag(dim)
mu1=rep(0,dim)
mu2=rep(0,dim)
n=256
tau=105
series1=createTimeSeries(mu1, mu2, sig1, n, tau)
sn=sin(2*pi*1:n/n)
series2=sn+series1
plot(series2[,2])
detectChangePoint(series2, showplot = TRUE)
#value less than 3 indicates no change-model is favored

speegled/cpbaywave documentation built on Feb. 19, 2024, 11:13 a.m.