Data-driven integration of norm-penalized mean-variance portfolios

TL;DR

This paper introduces a data-driven framework for optimizing norm-penalized mean-variance portfolios, improving robustness to estimation errors through a novel differentiable quadratic programming approach.

Contribution

It develops a neural network-based method to optimize convex combination of L1 and L2 penalties in mean-variance portfolios, with a new technique for derivative computation via implicit differentiation.

Findings

Enhanced portfolio stability in simulations

Reduced estimation error impact

Demonstrated benefits on US stocks and global futures data

Abstract

Mean-variance optimization (MVO) is known to be sensitive to estimation error in its inputs. Norm penalization of MVO programs is a regularization technique that can mitigate the adverse effects of estimation error. We augment the standard MVO program with a convex combination of parameterized $L_1$ and $L_2$-norm penalty functions. The resulting program is a parameterized quadratic program (QP) whose dual is a box-constrained QP. We make use of recent advances in neural network architecture for differentiable QPs and present a data-driven framework for optimizing parameterized norm-penalties to minimize the downstream MVO objective. We present a novel technique for computing the derivative of the optimal primal solution with respect to the parameterized $L_1$-norm penalty by implicit differentiation of the dual program. The primal solution is then recovered from the optimal dual variables. Historical simulations using US stocks and global futures data demonstrate the benefit of the data-driven optimization approach.