Building Classification Trees on Japanese Stock Groups Partitioned by Network Clustering

抄録

We built classification trees that provide sorting rules to reproduce the stock groups identified by hierarchical network clustering of Japanese stocks listed on the Tokyo Stock Exchange. The clustering method was based on a modularity maximization algorithm, which is frequently used for community detection in a network. We applied the clustering method to a correlation network in our previous research work. When building the classification trees, we tested various types of non-price data and selected effective variables with relative importance scores. The selection of variables proved to be consistent with a standard stock price model. Specifically, market capitalization and price book-value ratio were included as significantly important variables. Some other factors were also included as variables that clarify the properties of the Japanese stock market. The classification tree method can be applied to categorize stocks without a full set of continuous time series data into some groups, when our network clustering is difficult to be implemented. It can eventually contribute to improving risk measurement and management of stock portfolios.