Dynamic clustering of mutual funds based on the return series

Abstract

This paper proposes a method to analyze mutual funds using dynamic clustering based on the return series. For the time series data divided into some terms, the proposed method (1) converts the original high-dimensional data to two dimensional data using t-SNE for each term, (2) applies dynamic clustering using $x$-means for each terms, and (3) detects the cluster transitions using FBL-MONIC. This paper shows the experimental results for 29 Japanese mutual funds that track TOPIX, including four ETFs. The results indicate that there are three clusters at terms 1, 2, 3 and 4, and four clusters at term 5. We consider that the clusters are valid because one of the clusters consists of ETFs for all terms. FBL-MONIC could detect the transitions from a cluster at term 4 to a new cluster at term 5.