Why principled handling of uncertainty — not point-estimate optimisation — is the foundation of robust institutional portfolios.
Markowitz mean-variance optimisation, applied to historical sample estimates, is famously fragile. Tiny perturbations in expected returns produce wildly different portfolios. The problem is not optimisation; it is the pretence that we know the inputs.
The Bayesian fix
A Bayesian framework starts from a prior — typically the market portfolio or an equilibrium-implied set of returns — and updates it with the analyst’s views, weighted by the analyst’s confidence. The resulting posterior estimates produce portfolios that are interpretable, stable, and naturally regularised toward sensible baselines.
Black-Litterman as a special case
The Black-Litterman model, despite its age, remains a clean institutional implementation of this idea. Equilibrium returns implied by market caps form the prior; the user expresses absolute or relative views with explicit confidence; the posterior returns feed an optimiser. It is robust precisely because it does not pretend the user has perfect information.
Beyond returns: shrinkage on covariance
Sample covariance matrices are noisy at any reasonable lookback. Shrinkage methods — toward a constant-correlation target, a factor structure, or an industry-block model — are not optional refinements. They are the difference between a usable optimiser and a numerical curiosity.
FAQ
Is Black-Litterman still relevant?
Yes. The framework is provider-agnostic, transparent and understood by allocators. Modern variants extend it, but rarely replace it.