Clustering of Japanese stock returns by recursive modularity optimization for efficient portfolio diversification
Clustering of Japanese stock returns by recursive modularity optimization for efficient portfolio diversification
复制标题
DOI：
10.1093/comnet/cnu023
复制
发表时间：
2014-12-01
影响因子：
2.1
通讯作者：
Isogai, Takashi
中科院分区：
文献类型：
--
作者：
Isogai, Takashi
本文分析了日本股票收益率的高维相关结构，以找到一个比日本标准行业分类更面向数据和更灵活的分组，以实现有效的投资组合多样化，有助于更好的风险管理。利用广义自回归条件异方差模型对股票收益率进行滤波，分离收益率序列中的波动率。计算标准化收益率的相关矩阵，并基于相关矩阵建立股票收益率的无向网络。通过一系列模块化优化的递归谱聚类将股票收益分成若干组。我们开发了一种新的方法来控制递归聚类的过程，并确定最佳的组大小。将聚类结果与标准行业分类进行比较，通过统计检验探讨这两个群体之间的联系。标准部门分类法已被证明对确定某些类别是有效的;然而，即使在这种情况下，多个部门也被归入一个类别。有些团体与任何现有部门没有联系。我们进行随机投资组合模拟，以确认我们的分组是否有助于改善投资组合风险控制。仿真结果表明，在大多数情况下，基于新分组的样本投资组合的风险得到了较好的控制。本文提出的基于社区检测的方法也适用于其他厚尾金融资产收益率的聚类。
This paper analyses a high-dimensional correlation structure of Japanese stock returns to find a more data-oriented and flexible grouping than the Japan standard sector classification for efficient portfolio diversification that contributes to better risk management. The stock returns are filtered by a generalized autoregressive conditional heteroskedastic model to separate volatilities from return series. A correlation matrix of standardized returns is calculated, and an undirected network of the stock returns is built based on the correlation matrix. The stock returns are divided into several groups by a series of recursive spectral clustering with modularity optimization. We develop a new method to control the process of recursive clustering and determine the best group size. The clustering result is compared with the standard sector classification to explore how these two groups are linked by statistical tests. The standard sector classification is proved to be valid for identifying some groups; however, multiple sectors are included in a single group even in such cases. There are some groups that are not linked with any existing sector. We perform random portfolio simulations to confirm if our grouping can contribute to improving portfolio risk control. The simulation result shows that the risk of the sample portfolios based on the new grouping is better controlled in most cases. Our method based on community detection can be applicable for clustering other fat-tailed financial asset returns.