Stochastic regularization for the mean-variance allocation scheme

Abstract

Despite being based on sound principles, the original Markovitz portfolio allocation theory cannot produce sound allocations, and restrictions or modifications need to be imposed from outside the theory in order to obtain meaningful portfolios. This is unsatisfactory, and the reasons for this failure are discussed, in particular, the unavoidable small eigenvalues of the covariance. Within the original principles of risk minimization and return maximization, several modifications of the original theory are introduced. First, the strategic and tactical time horizons are separated. A base long-term allocation is chosen at the strategic time horizon, while the portfolio is optimized at the tactical time horizon using information from the price histories. Second, the tactical portfolio is financed by the strategic one, and a funding operator is introduced. The corresponding optimal allocation (without constraints) has one free parameter fixing the leverage. Third, the transaction costs are taken into account. This includes the current re-allocation cost, but crucially the expected costs of the next reallocation. This last term depends on the sensitivity of the allocation with respect to the covariance, and the expectation introduces another dependency on the (inverse) covariance. The new term regularizes the original minimization problem by modifying the lower part of the spectrum of the covariance, leading to meaningful portfolios. Without constraints, the final Lagrangian can be minimized analytically, with a solution that has a structure similar to the original Markovitz solution, but with the inverse covariance regularized by the expected transaction costs.

Keywords:

• Portfolio allocation
• Portfolio optimization
• Tactical asset allocation
• Covariance
• Markowitz theory

JEL classification:

• C61 (Optimization Techniques—Programming Models—DynamicAnalysis)
• G11 (Portfolio Choice)

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

1 Yes, the weights should decay regularly with the time lag, and not to be constant up to a cut-off l where they drop to zero. But the sliding window covariance is used broadly.