5.3.3 The Black–Litterman Formula

Basic Prerequisites: Posterior Inference and Reverse Optimization.

---

1. Intuition & Practical Objective

This page derives the core BL update: how views and the market prior combine into one posterior mean $\bar{\mu }$ and a posterior covariance $M$. The objective is a pair of formulas - the full conjugate form and the numerically-stable Master formula - plus the final unconstrained optimal-weights expression. If §02 gave you a sane starting place, this section gives you the blending algorithm.

The intuition is a leveraged weighted average. Your prior says “the market is right, $r\sim \mathcal{N}(\Pi ,\tau \Sigma )$.” Each view says “I think the market is wrong along direction $P_i$ by $Q_i$.” Where they meet is a precision-weighted compromise: pieces of the world you’re confident about (small $\Omega$) get honored almost fully; pieces the market calls (large prior uncertainty or weak view) get mostly ignored. The Master formula makes it explicit - the posterior is the prior plus a confidence-weighted correction proportional to the view residual $(Q-P\Pi )$.

---

2. Mathematical Ground Truth & Derivations

Setup. Prior on the (unknown, random) expected-return vector $r$:

$$r\sim \mathcal{N}(\Pi ,\tau \Sigma ).$$

Views, expressed as a noisy linear constraint on $r$ (Black–Litterman 1992, Idzorek eq. 5–6):

$$Pr=Q+\epsilon ,\epsilon \sim \mathcal{N}(0,\Omega ),$$

where $P$ is $K\times N$, $Q$ is $K$, $\Omega$ is $K\times K$. Each row of $P$ is a view: absolute view on asset $i$ has $P_{ki}=1$; relative view ”$i$ out-returns $j$ by $q$” has $P_{ki}=+1,P_{kj}=-1$ and $Q=q$.

Bayesian posterior (conjugate Gaussian). By the standard Normal–Normal update (the same machinery as Bayesian linear regression - ESL §3; see Bayes & Priors), the posterior is Normal with:

$$\bar{\mu }=[(\tau \Sigma {)}^{-1}+P^T{\Omega }^{-1}P{]}^{-1}[(\tau \Sigma {)}^{-1}\Pi +P^T{\Omega }^{-1}Q],$$ $$M=[(\tau \Sigma {)}^{-1}+P^T{\Omega }^{-1}P{]}^{-1}.$$

Here $(\tau \Sigma {)}^{-1}$ is the prior precision, $P^T{\Omega }^{-1}P$ the view precision projected into asset space, and the posterior precision is their sum. This is a precision-weighted average - exactly like a Bayesian ridge estimate.

The Master formula (numerically stable). Directly inverting $(\tau \Sigma {)}^{-1}$ invites numerical trouble. The Woodbury identity (Sherman–Morrison) transforms the mean into

$$\bar{\mu }=\Pi +\tau \Sigma P^T[P\tau \Sigma P^T+\Omega {]}^{-1}(Q-P\Pi ).$$

The structure is beautiful: $\Pi$ (market) plus the leverage $\tau \Sigma P^T$ times the Kalman gain $[P\tau \Sigma P^T+\Omega {]}^{-1}$ times the view residual $(Q-P\Pi )$. The gain matrix is only $K\times K$ (small) - that is why it is the workhorse implementation.

Final weights. Unconstrained MVO on $\bar{\mu }$ with risk-aversion $\delta$:

$$w^{\ast }=\frac{1}{\delta }{\Sigma }^{-1}\bar{\mu }.$$

Limits (shot through with sanity). If there are no views ($P=0$), $\bar{\mu }=\Pi$ and $w^{\ast }=\frac{1}{\delta }{\Sigma }^{-1}\Pi =w_{mkt}$ exactly. If a view has $\Omega \to \infty$ (no confidence), the gain $\to 0$ and again $\bar{\mu }\to \Pi$. Both identities are the first thing every implementation should test.

---

3. Computational Implementation - both forms, both limits

Verifies both formulas agree exactly, the zero-view identity recovers $w_{mkt}$, and the $\Omega \to \infty$ limit also returns to the market.

---

4. Failure Modes & First-Principles Breakdowns

1. Inverting $(\tau \Sigma {)}^{-1}$ directly. If $\Sigma$ is near-singular (near-collinear assets), $\tau \Sigma$ is too, and the full form amplifies noise. The Master formula never forms $(\tau \Sigma {)}^{-1}$ - use it.
2. **Confusing $M$ with $\Sigma$. $\bar{\mu }$’s covariance is $M$ (posterior estimate uncertainty), not the assets’ return covariance $\Sigma$. For portfolio risk (rather than return-estimate risk) you generally want $\Sigma +M$ combined - see 06 · Advanced Extensions.
3. Over-constraining views. If two views overlap the same assets ($P$ rows correlated), $P^T{\Omega }^{-1}P$ can be ill-conditioned; ensure $\Omega \ne 0$ and views are linearly independent.

---

5. References

• Black & Litterman (1992)
• Satchell & Scowcroft (2000)
• Idzorek (2005)
• Bayesian bridge: the Normal–Normal conjugate update is the same form as Bayesian linear regression

---

6. Connected Graph Bridges

• Base: Posterior Inference · Bayes & Priors · 02 · Reverse Optimization
• Continue: 04 · Views & Confidence · Index Hub