Find change points efficiently via rank-based change point detection

fastcpd_rank() and fastcpd.rank() detect change points using a rank-based, distribution-free cost. Each column is replaced by its global rank centred at zero, and change points in the mean of these centred ranks are detected with the existing PELT infrastructure. The result is fully deterministic and requires no bandwidth selection. The method is most powerful for location shifts; for scale-only or general distributional changes, fastcpd_kcp() is preferable.

Usage

fastcpd_rank(data, ...)

fastcpd.rank(data, ...)

Arguments

data

A numeric vector or a matrix with one row per observation.

...

Other arguments passed to fastcpd(), for example, segment_count.

Value

A fastcpd object.

See also

fastcpd()

Examples

set.seed(1)
x <- c(rt(200, df = 2), rt(200, df = 2, ncp = 5))
result_mean <- fastcpd_mean(x)
summary(result_mean)
#> 
#> Call:
#> fastcpd_mean(data = x)
#> 
#> Change points:
#> 332 333 
#> 
#> Cost values:
#> 3.352105 0 2.150836 
#> 
#> Parameters:
#>   segment 1 segment 2 segment 3
#> 1  3.436182  3047.762  8.251482
result_rank <- fastcpd_rank(x)
summary(result_rank)
#> 
#> Call:
#> fastcpd_rank(data = x)
#> 
#> Change points:
#> 200 
#> 
#> Cost values:
#> 102.8266 85.86544 
#> 
#> Parameters:
#>   segment 1 segment 2
#> 1    -94.44     94.44