Per August Jarval Moen: A grid-based methodology for online changepoint detection (and possible extensions to anomaly detection)

This talk introduces a novel grid-based methodology for online changepoint detection. The methodology enables the user to apply offline changepoint tests to sequentially observed data with low computational cost. Central to the methodology is a dynamically updating grid of candidate changepoints that "recycles" previously evaluated locations. For a specific class of test statistics, this approach ensures update and storage costs that scale logarithmically with sample size. We will explore the theoretical statistical performance of the methodology in two distinct settings: a Gaussian change-in-mean model and a Gaussian change-in-covariance model, demonstrating its near minimax-rate optimality in both cases. Lastly, we will discuss extensions of the methodology to anomaly detection settings.