CORRELATION SURPRISES FOR INVESTMENT GAINS: A MACHINE LEARNING APPROACH

Abstract

This study presents a comprehensive empirical analysis of new correlation surprise investment policies that can provide investors with superior risk-adjusted returns as compared to traditional buy-hold and classic fixed mix rules. We explore machine learning techniques including nearest neighbors classification, classification tree analysis and boosting/ tree ensemble in order to develop effective correlation derived investment models. We obtain empirical evidence that investment policies formed using the three machine learning frameworks, especially the nearest neighbors tactic, are effective when applied to both the DJIA and the S&P 500. In fact, the nearest neighbors tactic can boost the annualized return of a buy-hold portfolio of the DJIA from 7.9% to 22.5%, while also lowering the maximum draw-down over our 9 year investment horizon from 52% to 43%.