A1128
Title: Fast change point detection in R with sequential gradient descent
Authors: Xingchi Li - Texas A&M University (United States) [presenting]
Xianyang Zhang - Texas A\&M University (United States)
Abstract: Change point analysis seeks to identify shifts in data sequences by minimizing a cost function augmented by a penalty on the number of breakpoints. However, processing long series with complex models remains computationally demanding. Fastcpd, an R package that integrates the Pruned Exact Linear Time (PELT) framework, is introduced with sequential gradient descent to prune unnecessary evaluations and accelerate detection. Fastcpd accommodates a variety of scenarios, including mean and variance changes in univariate and multivariate time series, regression coefficient shifts, and structural breaks in ARMA, GARCH, and VAR processes, while allowing users to specify custom cost functions and penalties. Benchmarking on large datasets demonstrates that fastcpd achieves order-of-magnitude speed-ups over existing R tools, enabling scalable change point analysis without sacrificing accuracy. With its combination of flexibility, efficiency, and scalability, fastcpd provides researchers and practitioners with a powerful toolkit for detecting and localizing structural changes across diverse applications.
