Time Series Clustering: Pattern Recognition, Forecasting, and Amortized Inference

Overview

Time series clustering is a central task in statistics, with wide-ranging applications in demography, finance, biology, medicine, and other fields. The objective is to partition a collection of unlabeled time series into homogeneous groups such that similar series are assigned to the same cluster, while dissimilar ones are placed in different clusters. Traditionally, time series clustering has been used primarily for pattern recognition, that is, for detecting and summarizing patterns in collections of time series. In this context, the first part of the talk introduces a method for clustering locally stationary time series based on a dissimilarity measure relying on local quantile autocorrelation estimators. The method is evaluated in a simulation study, and its practical relevance is demonstrated using particulate matter data from KAUST. The second part presents a clustering framework designed to enhance forecasting by grouping time series according to predictive accuracy, yielding improved results across benchmark datasets. Finally, the talk discusses how amortized inference can address major challenges in time series clustering, such as algorithm selection, optimization limitations, and determination of the number of clusters, offering a promising unified solution to these complex issues.