Black-Litterman — Bayesian portfolio construction

Black-Litterman (1992) is the Bayesian solution to the Markowitz instability problem. Instead of plugging raw historical means into the optimiser and getting wild weights, BL starts with equilibrium-implied returns (the prior), incorporates the investor's subjective views (the likelihood), and produces a posterior that is the actual input to the mean-variance optimisation. The result: stable, intuitive portfolios that reflect the investor's information advantage where they have one and default to the market where they don't.

The two ingredients

Equilibrium returns Π (the prior)

Use the CAPM in reverse. The market portfolio is mean-variance efficient if and only if Π = λ Σ w_market — where λ is the market risk-aversion and w_market is the observed market-cap-weighted portfolio. This is reverse optimisation: instead of computing weights from expected returns, compute the expected returns that would make observed weights optimal.

Π = λ Σ w_market
λ = (μ_market - r_f) / σ²_market

Views Q and confidence Ω

P is a k × n 'view matrix'. Each row encodes one view as a linear combination of assets. Examples:

• Absolute view: 'Asset 1 will return 10%' — row is e_1 (unit vector), Q row is 0.10.
• Relative view: 'Asset 1 will outperform asset 2 by 3%' — row is e_1 - e_2, Q row is 0.03.
• Mixed view: 'Portfolio (50% asset 1, 50% asset 2) will outperform asset 3 by 2%' — row is (0.5, 0.5, -1), Q row is 0.02.

Ω is the k × k covariance of views — typically diagonal, with diagonal entry capturing the investor's confidence in each view (lower variance = higher confidence).

The Black-Litterman master formula

Π_BL = [(τΣ)⁻¹ + Pᵀ Ω⁻¹ P]⁻¹ [(τΣ)⁻¹ Π + Pᵀ Ω⁻¹ Q]

Posterior expected returns Π_BL are a precision-weighted average of the equilibrium prior Π and the view-implied returns. The scalar τ ∈ (0, 1) calibrates how much trust to place in the equilibrium prior relative to the data; common choices are 0.025 to 0.1.

Posterior covariance

Σ_BL = Σ + [(τΣ)⁻¹ + Pᵀ Ω⁻¹ P]⁻¹

Some BL variants use Σ_BL = Σ directly; the form above (He-Litterman 1999) adds parameter-uncertainty inflation. Either way, plug Π_BL and Σ_BL into the standard mean-variance optimiser to get the BL portfolio.

Why BL fixes the corner-solution disease

Standard Markowitz with sample means produces wild portfolios because sample means are extremely noisy. BL doesn't trust sample means; it trusts market caps (which reflect the consensus view) and the investor's own views (which the investor will own). The posterior μ tilts away from the market cap-weights only along the dimensions where views are stated. The result: BL portfolios deviate from market caps only along the directions of stated views, and only proportional to the confidence in those views.

Specifying Ω

Three common methods:

• He-Litterman (1999): Ω_jj = τ P_j Σ P_jᵀ. Each view's variance is proportional to its prior variance under τΣ.
• Idzorek (2005): specify a percentage confidence (e.g., 50%) per view; map back to Ω. More intuitive for non-quants.
• Direct: specify σ²_view per view explicitly. Used when you have explicit subjective uncertainty estimates.

BL workflow

1. Compute Σ (covariance of returns).
2. Compute Π = λ Σ w_market (equilibrium returns from market caps).
3. Encode views as (P, Q) and confidence Ω.
4. Apply the master formula to get Π_BL.
5. Run mean-variance optimisation with Π_BL (and Σ or Σ_BL).

BL in practice

• Most multi-strategy desks use a BL variant to combine quant models with discretionary PM views.
• Useful at multi-asset level: combine equilibrium-cap-weighted multi-asset benchmark with strategic asset allocation views.
• Idzorek's 'percentage confidence' parameterisation is industry-standard at endowments and SWFs.
• BL doesn't replace Markowitz — it pre-processes inputs for Markowitz.