Izzo, C; Medda, F; (2023) Higher order moments portfolio optimization via deep learning. Law and Economics Yearly Review , 12 (1) pp. 97-127.

Preview

Text Medda_articolo2023FinalIzzoMedda.pdf - Accepted Version Download (576kB) | Preview

Abstract

We analyse the problem of asset allocation with Deep Learning techniques via direct deterministic policy gradient methods applied to different reward formulations that depend on the desired degree of approximation of an investor's utility function, and therefore accounting for higher order moments. We compare against other deep learning and standard financial approaches. As we take an utilitarian view of the portfolio optimization problem, a contribution of this work is to compare the performances of Deep Learning Portfolio Optimization methods when using different objective functions depending on the desired degree of approximation of a given utility function. We test the different approaches via a recursive out-of-sample exercise on daily data from the 1st of January 2011 up until the 15th of March 2021. The investment set contains the sectoral decomposition of the SP500 plus the Bloomberg Barclays US Aggregate index. We find that including higher moments into the objective function helps improving performances, albeit the minimum-variance strategy combined with the Deep Dynamic Factor Model of Andreini et al. (2020) outperforms all the other methods over the sample analysed and in terms of the risk-adjusted metrics considered. We conclude this chapter by carrying out an analysis of the relation between the excess returns generated by selected strategies, and standard financial factors. We do so both by using linear regression methods and deep learning techniques together with Shapley values.

Download activity - last month

Export as CSVXMLJSON

Download activity - last 12 months

Export as CSVXMLJSON

Downloads by country - last 12 months

Export as JSONXMLCSV

Archive Staff Only

View Item