[Paper Review] Portfolio Optimization with Spectral Measures of Risk

Carlo Acerbi, Simonetti Prospero|ArXiv.org|Mar 29, 2002

Risk and Portfolio Optimization10 references30 citations

TL;DR

This paper extends the Pflug-Rockafellar-Uryasev methodology for Expected Shortfall optimization to general spectral measures of risk, showing that minimizing any spectral measure $ M_{ ho} $ is equivalent to minimizing a convex, piecewise-linear auxiliary function with additional parameters, enabling efficient linear programming solutions. The key insight is that minimizing a spectral measure inherently balances risk and return, making constrained Markowitz-style optimization equivalent to unconstrained minimization of an interpolated spectral measure.

ABSTRACT

We study Spectral Measures of Risk from the perspective of portfolio optimization. We derive exact results which extend to general Spectral Measures M_phi the Pflug--Rockafellar--Uryasev methodology for the minimization of alpha--Expected Shortfall. The minimization problem of a spectral measure is shown to be equivalent to the minimization of a suitable function which contains additional parameters, but displays analytical properties (piecewise linearity and convexity in all arguments, absence of sorting subroutines) which allow for efficient minimization procedures. In doing so we also reveal a new picture where the classical risk--reward problem a la Markowitz (minimizing risks with constrained returns or maximizing returns with constrained risks) is shown to coincide to the unconstrained optimization of a single suitable spectral measure. In other words, minimizing a spectral measure turns out to be already an optimization process itself, where risk minimization and returns maximization cannot be disentangled from each other.

Motivation & Objective

• To generalize the Pflug-Rockafellar-Uryasev method for Expected Shortfall minimization to arbitrary spectral measures of risk.
• To resolve the computational challenge of minimizing spectral measures due to their dependence on sorted scenario orderings.
• To reveal that minimizing a spectral measure inherently balances risk and return, eliminating the need for separate risk-return constraints.
• To demonstrate that constrained Markowitz-style optimization (minimize risk for fixed return or vice versa) is equivalent to unconstrained minimization of a single interpolated spectral measure.

Proposed method

• Propose a reformulation of spectral measure minimization using an auxiliary variable and a piecewise-linear, convex function $ \Gamma_{\hat{\phi}}(X, \psi_1) $, which avoids sorting subroutines.
• Use the spectral representation $ M_{\phi}(X) = -\int_0^1 \phi(p) F_X^{\leftarrow}(p) dp $ to define risk measures based on a risk aversion function $ \phi $.
• Introduce a linear interpolation between the spectral measure $ M_{\phi}(X) $ and the negative expected return $ -\mathrm{E}[X] $, forming $ M_{\hat{\phi}(\lambda)}(X) = (1-\lambda)M_{\phi}(X) - \lambda\mathrm{E}[X] $.
• Show that the minimization of $ M_{\hat{\phi}(\lambda)}(X) $ for $ \lambda \in [0,1] $ yields the efficient frontier of the classical Markowitz problem.
• Leverage the convexity and piecewise linearity of the reformulated function to enable efficient optimization via linear programming.
• Prove that non-optimal solutions of constrained problems correspond to minima of $ M_{\hat{\phi}(\lambda)}(X) $ only for $ \lambda \notin [0,1] $, ensuring only optimal portfolios are selected.

Experimental results

Research questions

• RQ1 Can the Pflug-Rockafellar-Uryasev method for Expected Shortfall minimization be generalized to arbitrary spectral measures of risk?
• RQ2 How can the non-analytic, sorting-dependent nature of spectral measure estimators be overcome for efficient optimization?
• RQ3 Is there a unified framework that unifies risk minimization and return maximization in portfolio optimization?
• RQ4 Can the classical Markowitz efficient frontier be recovered as an unconstrained minimization of a single spectral measure?
• RQ5 What is the role of the risk aversion function $ \phi $ in determining optimal portfolio allocations beyond standard deviation or Expected Shortfall?

Key findings

• The minimization of any spectral measure $ M_{\phi}(X) $ is equivalent to minimizing a convex, piecewise-linear function $ \Gamma_{\hat{\phi}}(X, \psi_1) $ with auxiliary variables, enabling efficient linear programming solutions.
• The constrained optimization problem of minimizing $ M_{\phi}(X) $ subject to a fixed expected return $ \mu $ is equivalent to unconstrained minimization of the interpolated spectral measure $ M_{\hat{\phi}(\lambda)}(X) = (1-\lambda)M_{\phi}(X) - \lambda\mathrm{E}[X] $ for $ \lambda \in [0,1] $.
• The efficient frontier in the $ (ES_\alpha, \mathrm{E}[X]) $ plane is recovered as the set of minimizers of $ M_{\hat{\phi}(\lambda)}(X) $ with $ \hat{\phi}(\lambda)(p) = \lambda + \frac{1-\lambda}{\alpha} \theta(\alpha - p) $.
• Minimizing a spectral measure inherently balances risk and return, making the separation of risk and return objectives in Markowitz-style optimization artificial and unnecessary.
• The method ensures that only optimal portfolios are selected, as non-optimal solutions of the constrained problem correspond to minima of $ M_{\hat{\phi}(\lambda)}(X) $ only for $ \lambda \notin [0,1] $, where the measure is no longer coherent.
• The reformulation allows for efficient optimization via standard linear programming techniques due to the piecewise linearity and convexity of the auxiliary function.