Portfolio Optimization based on Neural Networks Sensitivities from Assets Dynamics respect Common Drivers

TL;DR

This paper introduces a novel neural network-based framework for portfolio optimization that leverages sensitivities to common drivers, enhancing diversification and outperforming traditional methods across various markets.

Contribution

It presents the first use of neural network-approximated sensitivities and hierarchical clustering of these sensitivities for portfolio optimization, incorporating systematic and idiosyncratic risks.

Findings

Achieved over-performance in multiple market experiments.

Demonstrated effectiveness of hierarchical clustering on sensitivities.

Utilized public variables for maximum diversification.

Abstract

We present a framework for modeling asset and portfolio dynamics, incorporating this information into portfolio optimization. For this framework, we introduce the Commonality Principle, providing a solution for the optimal selection of portfolio drivers as the common drivers. Portfolio constituent dynamics are modeled by Partial Differential Equations, and solutions approximated with neural networks. Sensitivities with respect to the common drivers are obtained via Automatic Adjoint Differentiation. Information on asset dynamics is incorporated via sensitivities into portfolio optimization. Portfolio constituents are embedded into the space of sensitivities with respect to their common drivers, and a distance matrix in this space called the Sensitivity matrix is used to solve the convex optimization for diversification. The sensitivity matrix measures the similarity of the projections of portfolio constituents on a vector space formed by common drivers' returns and is used to optimize for diversification on both idiosyncratic and systematic risks while adding directionality and future behavior information via returns dynamics. For portfolio optimization, we perform hierarchical clustering on the sensitivity matrix. To the best of the author's knowledge, this is the first time that sensitivities' dynamics approximated with neural networks have been used for portfolio optimization. Secondly, that hierarchical clustering on a matrix of sensitivities is used to solve the convex optimization problem and incorporate the hierarchical information of these sensitivities. Thirdly, public and listed variables can be used to obtain maximum idiosyncratic and systematic diversification by means of the sensitivity space with respect to optimal portfolio drivers. We reach over-performance in many experiments with respect to all other out-of-sample methods for different markets and real datasets.