[Paper Review] Bayesian mean-variance analysis: Optimal portfolio selection under parameter uncertainty

David Bauder, Taras Bodnar+2 more|arXiv (Cornell University)|Mar 9, 2018

Risk and Portfolio Optimization54 references42 citations

TL;DR

This paper proposes a Bayesian mean-variance approach that formulates optimal portfolio selection using the posterior predictive distribution, eliminating dependence on unknown parameters by relying solely on historical data. By deriving a stochastic representation of the predictive distribution, the method yields a Bayesian efficient frontier that outperforms the classical sample efficient frontier, reducing overoptimism and enabling credible intervals for future portfolio returns.

ABSTRACT

The paper solves the problem of optimal portfolio choice when the parameters of the asset returns distribution, like the mean vector and the covariance matrix are unknown and have to be estimated by using historical data of the asset returns. The new approach employs the Bayesian posterior predictive distribution which is the distribution of the future realization of the asset returns given the observable sample. The parameters of the posterior predictive distributions are functions of the observed data values and, consequently, the solution of the optimization problem is expressed in terms of data only and does not depend on unknown quantities. In contrast, the optimization problem of the traditional approach is based on unknown quantities which are estimated in the second step leading to a suboptimal solution. We also derive a very useful stochastic representation of the posterior predictive distribution whose application leads not only to the solution of the considered optimization problem, but provides the posterior predictive distribution of the optimal portfolio return used to construct a prediction interval. A Bayesian efficient frontier, a set of optimal portfolios obtained by employing the posterior predictive distribution, is constructed as well. Theoretically and using real data we show that the Bayesian efficient frontier outperforms the sample efficient frontier, a common estimator of the set of optimal portfolios known to be overoptimistic.

Motivation & Objective

• Address the well-known overoptimism of the classical sample efficient frontier due to parameter estimation error.
• Overcome the suboptimal nature of two-step mean-variance optimization, where estimated parameters are used in a second step to compute portfolio weights.
• Develop a fully Bayesian approach that incorporates parameter uncertainty from the outset by using the posterior predictive distribution.
• Construct a Bayesian efficient frontier that reflects both estimation and model uncertainty, improving out-of-sample performance.
• Provide a stochastic representation enabling prediction intervals for optimal portfolio returns, enhancing risk assessment and decision-making.

Proposed method

• Formulate the portfolio optimization problem using the posterior predictive distribution of future asset returns, which integrates over the uncertainty in the mean vector and covariance matrix.
• Derive a stochastic representation of the posterior predictive distribution using a multivariate t-distribution, based on the joint posterior of the mean and covariance matrix under Jeffreys’ non-informative prior.
• Use the stochastic representation to analytically derive the posterior mean and variance of the portfolio return, which are functions of observed data only.
• Construct the Bayesian efficient frontier by maximizing the posterior mean-variance utility function across all feasible portfolios.
• Apply the derived stochastic representation to generate prediction intervals for optimal portfolio returns, accounting for both estimation and model uncertainty.
• Employ the inverse Wishart and multivariate t-distributions to model the posterior distributions of the covariance matrix and mean vector, respectively.

Experimental results

Research questions

• RQ1 How can parameter uncertainty in mean and covariance estimates be properly incorporated into mean-variance portfolio optimization to avoid suboptimal results?
• RQ2 Can a fully Bayesian approach that uses the posterior predictive distribution yield a more robust and less overconfident efficient frontier than the classical sample efficient frontier?
• RQ3 What is the analytical structure of the posterior predictive distribution of portfolio returns under non-informative priors, and how can it be used for portfolio construction?
• RQ4 How do prediction intervals for optimal portfolio returns derived from the posterior predictive distribution compare to those from classical methods in terms of coverage and width?
• RQ5 To what extent does increasing sample size reduce estimation risk in the Bayesian framework, and does this offset the increased economic risk in riskier portfolios?

Key findings

• The Bayesian efficient frontier, derived from the posterior predictive distribution, systematically outperforms the classical sample efficient frontier by reducing overoptimism and improving out-of-sample performance.
• The posterior predictive distribution of portfolio returns is analytically derived as a scaled non-central t-distribution, enabling exact computation of moments and prediction intervals.
• Prediction intervals for optimal portfolio returns are constructed using the stochastic representation, showing that credible intervals widen with risk, even as sample size increases, reflecting persistent economic risk.
• The Bayesian approach yields optimal portfolio weights that depend only on observed historical data, eliminating reliance on estimated parameters and ensuring a fully implementable solution.
• The posterior mean of the portfolio return is equal to the sample mean return, but the posterior variance is larger than the classical sample variance, reflecting the inclusion of estimation uncertainty.
• Empirical results confirm that credible intervals for portfolio returns do not cover negative values at the 95% confidence level, indicating high confidence in positive expected returns under the Bayesian framework.