NSP wrapper — Narrowest Significance Pursuit — nsp

Wraps the nsp package (Fryzlewicz 2024). NSP inverts the usual framing of post-selection inference: rather than estimating changepoint locations and then asking whether they are real, it returns a set of intervals, each of which contains at least one changepoint, with the guarantee holding globally across all intervals simultaneously at level alpha. The guarantee is exact and finite-sample, and the self-normalised and autoregressive variants keep it under heavy tails, heteroscedasticity and serial dependence.

Usage

nsp_wrapper(
  x,
  alpha = 0.1,
  variant = c("poly", "selfnorm", "ar", "tvreg"),
  change_in = c("mean", "slope"),
  deg = NULL,
  M = 1000,
  covariates = NULL,
  ord = 1,
  seed = NULL,
  ...
)

Arguments

x

A numeric vector.

alpha

Global significance level: with probability at least \(1 - \alpha\), every returned interval contains a changepoint. Defaults to 0.1.

variant

Which NSP procedure to run:

"poly": (default) nsp::nsp_poly() — piecewise polynomial signal, Gaussian noise of constant variance.
"selfnorm": nsp::nsp_poly_selfnorm() — self-normalised, for heavy tails and heteroscedasticity. Slower.
"ar": nsp::nsp_poly_ar() — autoregressive noise of order ord.
"tvreg": nsp::nsp_tvreg() — a general linear model whose coefficients change; requires covariates.

change_in

"mean" (a piecewise-constant signal, deg = 0) or "slope" (piecewise linear, deg = 1). Ignored when deg is given explicitly, and when variant = "tvreg" (which takes its model from covariates).

deg

Degree of the piecewise polynomial. Derived from change_in when NULL.

M

Number of intervals drawn. Defaults to 1000; the engine's own default.

covariates

A design matrix for variant = "tvreg", with one row per observation.

ord

AR order for variant = "ar". Defaults to 1.

seed

Optional seed. NSP draws random intervals, so a run is reproducible only with one.

...

Additional arguments passed to the underlying nsp function.

Value

A ggcpt object with a populated regions slot.

What `cp` means here, and what it does not

NSP produces no point estimates. This wrapper still fills the cp column — with the midpoint of each interval — because every downstream consumer in the package (augment(), cpt_metrics(), cpt_consensus(), autoplot()) is built on that column, and a result with an empty cp would silently score as "found nothing". The midpoint is not an estimate of the changepoint location and must not be reported as one: the interval is the inferential object. The result therefore

carries the intervals in a regions slot, read with cpt_regions();
marks itself, so print() says the cp column is a midpoint and autoplot() shades the bands by default;
adds a cp_source column reading "region_midpoint" to the changepoints tibble.

References

Fryzlewicz P (2024). “Narrowest significance pursuit: inference for multiple change-points in linear models.” Journal of the American Statistical Association, 119(546), 1633–1646. doi:10.1080/01621459.2023.2211733 .

Other changepoint engines: bcp_wrapper(), beast_wrapper(), bfast_wrapper(), binsegrcpp_wrapper(), bocpd_wrapper(), cpm_wrapper(), cpop_wrapper(), cpt_wrapper(), decafs_wrapper(), ecp_wrapper(), envcpt_wrapper(), esac_wrapper(), fabisearch_wrapper(), fastcpd_wrapper(), fcov_wrapper(), fmean_wrapper(), fpop_wrapper(), geomcp_wrapper(), hdcov_wrapper(), hdreg_wrapper(), idetect_wrapper(), inspect_wrapper(), kcp_wrapper(), kwc_wrapper(), mcp_wrapper(), mosum_wrapper(), network_wrapper(), not_wrapper(), npmojo_wrapper(), ocd_wrapper(), pilliat_wrapper(), segmented_wrapper(), smuce_wrapper(), sn_wrapper(), strucchange_wrapper(), taylor_wrapper(), tguh_wrapper(), trend_wrapper(), var_wrapper(), wbs2_wrapper(), wbs_wrapper(), wbsts_wrapper()

Examples

set.seed(2026)
x <- c(rnorm(100), rnorm(100, 4))
fit <- nsp_wrapper(x, M = 100, seed = 1)
cpt_regions(fit)
#> # A tibble: 1 × 4
#>   start   end length value
#>   <int> <int>  <int> <dbl>
#> 1    96   105     10  4.54
ggplot2::autoplot(fit)

NSP wrapper — Narrowest Significance Pursuit