An arXiv preprint dated 28 Aug 2026 describes SCAN, a statistical method designed to find multiple points where the distribution of a long, serially dependent univariate time series changes. Across simulations, its strongest relative results came on moderately long to very long series, although the picture was less favorable in long-memory scenarios.
How the detector works
SCAN sets local decision thresholds with a tapered block bootstrap and applies a Bonferroni correction across those local tests. The adjusted level is the nominal level divided by the number of local tests, a design intended to control family-wise error without assuming independence.
It then runs detectors at different window sizes, clusters nearby detections and keeps clusters that reach a user-specified support threshold. Within each retained cluster, it selects the most-supported location.
A result with a narrow boundary
One theoretical result gives SCAN a more familiar shape. Under a single pure mean shift, its population SWAL localization criterion equals the corresponding CUSUM-type criterion. The equivalence is limited to the stated model, which assumes common mean-zero noise and finite first moments.
The theoretical section also states consistency for both the estimated number and the relative locations of change-points under its stated assumptions. The supplied analysis notes that uniform bootstrap validity over a growing collection of dependent local tests was not established.
The pattern in the simulations
The tests covered mean changes and joint mean-and-variance changes at series lengths of 500, 1,000, 5,000, 10,000, 20,000, 50,000, 100,000 and 1,000,000 observations. Each scenario used 1,000 replications, with dependence structures ranging from independence to long-range dependence.
For mean changes, SCAN was competitive on short series and showed its strongest relative performance on moderately long to very long series. The authors report generally higher detection accuracy than fixed-penalty competitors, with similar F1 patterns.
For joint mean-and-variance changes, SCAN had its strongest relative covering performance on moderately long to very long series. It generally reported higher detection and localization accuracy than BIC-penalty competitors, while results were more mixed on short series.
A clear boundary appeared under ARFIMA long-memory dependence. SCAN performed less favorably and became more sensitive to the ensemble voting threshold. The paper identifies its exponential alpha-mixing calibration assumption as the stated explanation.
Where the method met real data
In an application to HASC Series 2 from individual 671-a, the data contained approximately 40,000 observations and 38 annotated activity transitions. Using a matching tolerance of 120 observations, SCAN selected 37 change-points and achieved an F1-score of 0.80.
The reported HASC covering metric was 0.8360, and most detected change-points were described as closely aligned with annotated activity transitions. This was an annotation-based descriptive evaluation rather than an independently adjudicated outcome.
The Bitcoin example used 81,079 hourly BTC-USD observations covering 3,378 days. SCAN detected changes on 40 distinct dates, and 33 of those dates were linked to relevant Reuters reports.
Those report links offer potentially interpretable event-date alignment, but they do not validate detection accuracy or show that the reports caused the changes. The supplied analysis notes that Bitcoin has no stated gold-standard change-point labels.
What remains uncertain
Taken together, the results point to a tool aimed at long, dependent univariate streams, with stronger reported results in longer settings and a clear warning under long-memory dependence. The simulations did not make SCAN a universal winner: short-series outcomes were mixed, and some competing methods performed similarly or better in selected settings.
The supplied document does not report a funding source. The authors report no competing interests and disclose using ChatGPT-5 for proofreading and language enhancement, followed by author review and editing.
Paper data and sources
Original title: SCAN: Sequentially Detecting Change-points via Adaptive Nonparametric Inference
Authors: Ashoka Prabashwara, Patricia Menéndez, Liam Hodgkinson, Stuart Lee
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: Not available
Original paper · Full text