The Black-Litterman Model

Black and Litterman [@black1992] addressed the brittleness of plug-in mean-variance by combining a market-implied prior on returns with investor views.

Reverse optimization: the prior

Assume the market portfolio \(w_{\mathrm{mkt}}\) is the optimum of an unconstrained mean-variance problem with risk aversion \(\delta\). Inverting the first-order condition gives the equilibrium implied returns:

\[ \Pi \;=\; \delta\, \Sigma\, w_{\mathrm{mkt}}. \tag{1} \]

The prior on the true (unobserved) mean \(\mu\) is

\[ \mu \;\sim\; \mathcal{N}\bigl(\Pi,\; \tau \Sigma\bigr), \tag{2} \]

where \(\tau\) is a small scalar reflecting confidence in the equilibrium (typical values \(0.01 \le \tau \le 0.05\)).

Encoding views

A view is a linear statement about \(\mu\):

\[ P \mu \;=\; q \;+\; \varepsilon, \qquad \varepsilon \sim \mathcal{N}(0,\, \Omega), \tag{3} \]

where \(P \in \mathbb{R}^{k \times n}\) picks linear combinations of assets, \(q \in \mathbb{R}^k\) is the view magnitude, and \(\Omega \in \mathbb{R}^{k \times k}\) is the view uncertainty (diagonal in practice).

Posterior

Combining (2) and (3) by standard Gaussian conjugacy gives the posterior mean of \(\mu\):

\[ \bar\mu \;=\; \Bigl[(\tau \Sigma)^{-1} + P^\top \Omega^{-1} P\Bigr]^{-1} \Bigl[(\tau \Sigma)^{-1} \Pi + P^\top \Omega^{-1} q\Bigr], \tag{4} \]

and posterior covariance

\[ M \;=\; \bigl[(\tau \Sigma)^{-1} + P^\top \Omega^{-1} P\bigr]^{-1}. \tag{5} \]

The Black-Litterman portfolio plugs \((\bar\mu,\; \Sigma + M)\) into the mean-variance optimizer of Mean-Variance.

Setting \(\Omega\): the Idzorek convention

Idzorek [@idzorek2005] specifies \(\Omega\) by confidence levels \(c_k \in [0, 1]\) rather than variances. For view \(k\) with row \(p_k\), set

\[ \Omega_{kk} \;=\; \frac{1 - c_k}{c_k} \cdot p_k\, (\tau \Sigma)\, p_k^\top. \tag{6} \]

This recovers the intuitive limits: \(c_k \to 0\) ignores the view, \(c_k \to 1\) enforces it exactly.

He-Litterman convention

He and Litterman [@he1999] use a simpler diagonal form:

\[ \Omega \;=\; \mathrm{diag}\bigl(P\, (\tau \Sigma)\, P^\top\bigr), \tag{7} \]

which implicitly assigns each view a confidence consistent with its projection of the prior covariance. The He-Litterman 1999 validation page reproduces the eight-country example using this convention.

Practical notes

• Always shrink \(\Sigma\) before reverse-optimization; the equilibrium weights are extremely sensitive to small eigenvalues.
• For a fully invested optimization, use \(\delta\) implied by the historical market Sharpe rather than the textbook \(\delta = 2.5\).
• The library exposes BlackLitterman(prior=..., views=..., omega="idzorek") with both conventions selectable.

References

[@black1992]; [@he1999]; [@idzorek2005]. See Citations.