Title: Machine learning, economic regimes and portfolio optimisation

Authors: John M. Mulvey; Han Hao; Nongchao Li

Addresses: Bendheim Center for Finance, Operations Research and Financial Engineering Department, Princeton University, USA ' Bendheim Center for Finance, Operations Research and Financial Engineering Department, Princeton University, USA ' Bendheim Center for Finance, Operations Research and Financial Engineering Department, Princeton University, USA

Abstract: In portfolio models, the depiction of future outcomes depends upon a representative accounting of economic conditions. There is much evidence that crash periods display much different patterns than normal markets, suggesting that forecasting models ought to be based on multiple regimes. We apply two techniques from machine learning in our empirical study to improve robustness: 1) trend-filtering - to distinguish regimes possessing relatively homogeneous patterns; 2) a shrinkage/cross validation approach within a factor analysis of performance. A scenario-based portfolio model is proposed and designed to address multiple regimes. The worst-case events are well described within the framework, as compared with mean-variance Markowitz models that treat equally all historical performance.

Keywords: portfolio models; asset allocation; economic regimes; machine learning; classification; factor investing; worst-case events.

DOI: 10.1504/IJFERM.2018.094043

International Journal of Financial Engineering and Risk Management, 2018 Vol.2 No.4, pp.260 - 282

Available online: 03 Aug 2018 *

Full-text access for editors Access for subscribers Purchase this article Comment on this article