Mathematical Foundation for Ensemble Machine Learning and Ensemble Portfolio Analysis

Abstract

In ensemble machine learning we combine decisions of experts to derive at a decision that is better than the individual ones. The process of combining these decisions can be as simple as majority voting or simple averaging or it can be more complicated and involve multiple steps. In this paper we consider the application of ensemble machine learning to the problem of constructing portfolios from individual decisions of multiple experts. We will compare the performance of portfolios constructed by simple averaging and by a novel multistage decision algorithm. This new algorithm constructs a portfolio from subsets of stocks in individual portfolios. Compared to these individual portfolios and a portfolio constructed by simple averaging, the portfolios constructed by proposed method could result in higher annualized return and a modest increase in volatility. We provide extensive numerical comparisons on the viability of the new approach.