Dimension Reduction on Statistical Moment Term Structure for Portfolio Diversification

Abstract

Segmentation, the process of dividing a target market into sub-sections, is widely used in the equity markets to help investors limit or increase exposure to a set of stocks with specific common risk and economic characteristics, which underlies the philosophy of diversification. Traditional diversification can be achieved by investing in uncorrelated markets segments which are defined by industrial classifications. The latter is a naive methodology that can offer little diversification to investors, by virtue of being a product of the strict qualitative methodology in defining the industrial sectors. This paper explores a new methodology using dimensionality reduction to select stocks and partition the market into based on higher order statistical moments. These new sectors are chosen to maximize the explanatory power of stocks on the variance of statistical structure and are thus poised to give investors greater risk segmentation. Based on a risk adjusted performance, we find that the portfolios generally outperform both the market and a naive industry sector portfolio. Finally, we note that there is local stable quantile structure in the moments of our portfolios, giving proof that our segmentation may be viable long term segmentation and diversification strategy.