The gridcp package, available on PyPI, makes offline (fixed-sized) changepoint tests easy to apply in the online (streaming) setting.
While an abundance of offline changepoint tests and methods exist, they are typically challenging to apply directly in streaming scenarios, where observations must be processed one at a time. This is because the per-observation processing cost typically grow at least linearly with sample size. The recently proposed grid-based methodology of Moen (2026) alleviates this by evaluating offline test statistics over a sparse geometric grid of split points, spaced increasingly far apart further into the past.
Our new preprint documents gridcp, which implements this methodology behind a single uniform interface. It provides nine Numba-accelerated pre-implemented changepoint tests, covering changes in mean, variance, covariance and regression coefficients as well as nonparametric and exponential-family likelihood-ratio tests. Users may also specify custom changepoint tests, which the package handles just as easily as the built-in ones. gridcp further offers Monte Carlo routines for calibrating detection thresholds to a target false alarm probability or average run length, including data-driven variants.
gridcp preprint: arXiv:2608.18695
Methodology preprint (Moen, 2026): arXiv:2504.09573