A Dynamic yet Stable Time-Varying Clustering Method

I propose a method to cluster multivariate panel data stably, such that switches between clusters are minimized over time. This results in clusters that are dynamic, varying in size, location and number, but also inert, meaning that those changes happen slowly over time. This method leverages on the past clustering while also producing dynamic compositions. The core of the algorithm consists in shrinking the distance between a point and its previous cluster and applying a cross-sectional clustering method. I discuss methods for validation and visualization in a simulation setting, and also in an empirical application using insurance firms' balance sheet data from WRDS. There, I find that dynamic clusters complement industry classification standards. Finally, I propose extensions of the algorithm to accommodate a wider class of cross-sectional clustering methods.