66th ISI World Statistics Congress

66th ISI World Statistics Congress

Changepoint Detection in Hierarchical Additive Time Series

Conference

66th ISI World Statistics Congress

Format: WSC 2027: IPS Abstract

Abstract

Hierarchical Additive time series occur in many settings, for example, businesses record sales of individual products, product categories, stores, regions, and the company as a whole. Changepoints can occur at any or all of these levels.
Changepoint detection in hierarchical time series is complicated by the additive aggregation relationships linking series across hierarchical levels. Structural changes originating in component series may be reinforced, attenuated, or cancelled through aggregation, so that the evidence for a changepoint need not be equally strong at a node and its ancestors.
Identifying which levels exhibit changes can help to discern, for example, whether promotions increase sales for a category overall, or just displace purchases from one product line to another. Similarly, small changes across many product lines may be indistinguishable individually, but together are detectable.
Multivariate changepoint detection doesn't account for the inherent additivity structure across the time series and so new approaches for respecting the hierarchy need to be developed. In this presentation, we will discuss recent research for detecting changepoints in hierarchical additive time series.