Bayesian portfolio optimisation: priors, posteriors, and shrinkage in mean-variance inputs
Mean-variance optimisation is famously sensitive to its expected-return inputs: small changes in assumed means produce large changes in optimal weights, and the resulting portfolios are often concentrated and counter-intuitive. The Bayesian framework addresses this by combining prior beliefs about expected returns with the sample evidence, producing posterior estimates that are more stable and that translate into more reasonable portfolios.
What Bayesian portfolio optimisation is
The Bayesian approach treats expected returns and the covariance matrix as unknown quantities about which the investor has prior beliefs. The prior is updated by Bayes' rule using the sample of historical returns, producing a posterior distribution that incorporates both sources of information. The optimisation is then run using the posterior means and covariances as inputs.
The strength of the prior controls how much the posterior moves with the data. A strong prior produces a posterior that barely shifts even with substantial sample evidence; a weak prior collapses to the standard sample-based estimates. Calibrating the prior is the modeller's central task, and different choices of prior produce different optimal portfolios for the same data.
How it works
The Black-Litterman model (Black & Litterman, 1992) is the most-cited Bayesian construction in investing. It uses the market-implied expected returns as the prior—the equilibrium-weighted expected returns that would justify the observed market portfolio—and combines those with the investor's specific views to produce posterior expected returns. The framework allows the investor to express views with different confidence levels, which the model converts into appropriate weights on the prior versus the views.
Other Bayesian approaches include Stein-James shrinkage (which shrinks sample means towards the grand mean), Ledoit-Wolf shrinkage of the covariance matrix (which shrinks the sample covariance towards a structured target), and full hierarchical Bayesian models that estimate priors from cross-sectional patterns. All share the goal of stabilising the optimisation inputs without discarding the sample evidence entirely.
What the evidence shows
Empirical studies of Bayesian portfolio construction (DeMiguel, Garlappi & Uppal, 2009; Tu & Zhou, 2011) consistently find that Bayesian estimators produce more stable out-of-sample portfolios than naive mean-variance optimisation. The improvement is largest when the sample is small or noisy and shrinks when long, clean histories are available.
The Black-Litterman framework specifically has been adopted widely by institutional asset allocators, including some of the largest sovereign wealth funds and pension plans. The framework's appeal is partly empirical (it produces reasonable portfolios) and partly behavioural (it gives investors a structured way to incorporate their views without producing unreasonable concentration).
Limitations and trade-offs
The Bayesian framework requires a prior, and the choice of prior is genuinely subjective. The market-implied prior in Black-Litterman is a defensible choice but rests on the assumption that current market weights reflect equilibrium expectations—itself a strong assumption. Other priors (e.g., minimum-variance, equal-weight) make different implicit claims.
Bayesian optimisation also does not solve the fundamental problem that future expected returns are unknown. It substitutes a more stable estimation procedure for a less stable one, which improves out-of-sample performance, but the basic difficulty of estimating means from finite return histories remains.
Bayesian portfolio optimisation in pfolio
Bayesian portfolio optimisation is not implemented in pfolio's pre-built portfolios. The platform's systematic methodology uses a momentum-based ranking rather than mean-variance inputs that would require Bayesian shrinkage; the construction methodology is documented at how we build portfolios.
Related articles
Disclaimer
Get started now
