Abstract

Many real-world time series are nonstationary and exhibit structural changes due to shifts in the underlying data-generating mechanism. Overlooking such changes and assuming a constant data-generating mechanism may result in invalid statistical inference. While change point detection is an important tool for identifying and localising changes in time series, most existing methods assume complete observations and cannot be directly applied to incomplete data. A nonparametric method for change point detection under missingness is introduced by extending and reweighting a kernel-based framework to accommodate incomplete observations without requiring imputation. The proposed method retains the ability to detect changes in the distribution of the underlying process in the presence of missing values and does not require independence of the observations. Theoretical guarantees are established for the proposed approach, and the method's utility is demonstrated on an environmental data application.