Black-Litterman (1992) Global Portfolio Optimization

Summary

This is the founding statement of the Black-Litterman model for global asset allocation. It diagnoses why standard mean-variance optimization is impractical — it is hyper-sensitive to the expected-return inputs and produces extreme, unbalanced long/short portfolios — and fixes it by anchoring expected returns to a neutral equilibrium reference: the CAPM risk premiums implied by market-capitalization weights (reverse optimization). Investor views, absolute or relative and each with a stated confidence, are then blended with this equilibrium prior in a Bayesian fashion, yielding posterior expected returns that produce well-behaved portfolios which tilt the market portfolio toward favored assets in proportion to the strength of the views. The setting is global equities, bonds, and currencies with currency hedging.

Key Claims

Mean-variance is input-sensitive. Standard optimization "is extremely sensitive to the expected return assumptions the investor must provide," generating unreasonable portfolios with large long/short positions unless artificially constrained.
Equilibrium as the neutral prior. The CAPM equilibrium risk premiums — the excess returns that clear supply and demand for global assets given market-cap weights — are a "center of gravity" for expected returns and the correct neutral reference when the investor has no views. Historical-average and equal-mean alternatives give poor, non-neutral results.
Reverse optimization. Equilibrium expected excess returns are recovered from market-cap weights and the covariance matrix, $\Pi = \delta\,\Sigma\,w_{\text{mkt}}$ (implied risk premiums), rather than estimated from historical means.
Views blended by mixed estimation. Expected returns are modeled as an equilibrium value plus uncertainty; investor views (of the form "I expect $P\,r = Q$ ," absolute or relative, with confidence $\Omega$ ) are combined with the prior using the Theil (1971) mixed-estimation / Bayesian updating logic, with $\tau$ scaling the uncertainty in the equilibrium mean.
Well-behaved, confidence-weighted portfolios. The posterior expected returns deviate from equilibrium only in the directions of the investor's views, and only as far as their confidence warrants, so the optimal portfolio is the market portfolio tilted toward favored assets — balanced, intuitive, and stable.
Relative views and partial information. The investor may state as few or as many views as desired, including purely relative views (asset A will outperform asset B), which standard models cannot accommodate without a full set of absolute return forecasts.

Concepts Introduced or Extended

Entities Mentioned

Quotes

"Equilibrium risk premiums provide a center of gravity for expected returns."

"Rather than requiring the investor to have a view about the absolute return on every asset and currency, our approach allows the investor to specify as many or as few views as he wishes… and can specify a degree of confidence about each view."

My Take

The model's durability comes from turning a numerical pathology into a modeling principle: the reason unconstrained mean-variance blows up is that it treats noisy return estimates as certain, so Black-Litterman replaces the point estimate with a prior (equilibrium) and updates it only where the investor has information. That is exactly the shrinkage/mixed-estimation idea Litterman had already deployed for VAR forecasting in the Minnesota prior — a diffuse-data problem tamed by an economically motivated prior — reused here for the cross-section of returns. It is a complementary answer to the estimation-risk literature: that work quantifies how badly sample means mislead, while Black-Litterman supplies a constructive prior to lean on instead. The practical friction is the free parameters — $\tau$ and the view-confidence $\Omega$ have no canonical calibration, and results can move with them — which the original article treats more by intuition than by formal rule.