5.3.1 Black–Litterman from Zero

Basic Prerequisites: Mean–Variance from Zero.

---

1. Intuition & Practical Objective

This page builds the why of Black–Litterman with no portfolio-optimization experience needed beyond the one-line spec of Markowitz. The objective is one idea: when you hand a mean-variance optimizer your best-guess expected returns, it doesn’t optimize - it amplifies your noise into extreme bets you never meant to make. Black–Litterman’s cure is to stop guessing and start from the one portfolio nobody argues with: the market itself.

Start with the dumbest question: what expected returns should I feed the optimizer? Markowitz tells you $w^{\ast }=\frac{1}{\delta }{\Sigma }^{-1}(\mu -r_{f}1)$. The problem: you do not know $\mu$. Sample means are notoriously noisy (estimation error scales like $\sigma /\sqrt{T}$). What the optimizer does with that noise is the crime: because $w^{\ast }$ is linear in ${\Sigma }^{-1}$, any error in $\mu$ is multiplied by the inverse covariance, whose tiny-eigenvalue directions blow up. The result is weights like $[+0.90,-0.30,+0.93]$ that look like a bet on apocalypse.

Three intuitions (the “aha”s):

1. The market is the neutral answer. If you have no view on anything, the rational thing is to own the world - the market-cap portfolio. So any sensible starting point should reproduce $w_{mkt}$ when you have no information. Let the optimizer endorse the market by default and only deviate where you know something.
2. Expected returns can be derived, not guessed. Rather than estimating $\mu$ with noisy statistics, invert Markowitz on the known market portfolio. Ask: “what expected returns would justify holding $w_{mkt}$?” That reverses the formula into the implied returns $\Pi =\delta \Sigma w_{mkt}$ - a stable, supply-driven prior.
3. Your opinion enters as a gentle tilt, not a takeover. You rarely have a complete forecast for all $N$ assets - usually a handful of views (“tech beats utilities by 4%”). BL encodes only those views and blends them with the prior by confidence, leaving untouched assets at their equilibrium weight. Noise that MVO would scatter everywhere now has nowhere to hide.

---

2. Mathematical Ground Truth & Derivations

The failure of naive MVO. Feed the optimizer arbitrary sample means ${\mu }_s$; the weights are

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

The sensitivity to a perturbation $d\mu$ is $d{w}^{\ast }=\frac{1}{\delta }{\Sigma }^{-1}d\mu$. Because ${\Sigma }^{-1}$‘s eigenvalues are $1/{\lambda }_i$, small inputs along low-variance directions get amplified. Best & Grauer (1991) prove formally that a one-percent change in a single asset’s mean can drive that asset to its max short in a 100-asset problem. This is “garbage in, garbage out” with compounding interest.

The BL prior (sketch). The market-cap portfolio $w_{mkt}$ is assumed (reverse) optimal. Matching the MVO first-order condition gives the implied returns

$$\Pi =\delta \Sigma w_{mkt},$$

so that $\frac{1}{\delta }{\Sigma }^{-1}\Pi =w_{mkt}$ exactly. Rather than a point estimate of $\mu$, BL treats returns as random centered on $\Pi$:

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

where $\tau \Sigma$ is the prior’s covariance - the uncertainty in our knowledge of the mean, scaled down from $\Sigma$ by $\tau \in (0,1)$. This is the Bayesian prior; §02 derives where $\Pi$ comes from, §03 does the update.

---

3. Computational Implementation - MVO noise vs. BL stability

Direct comparison on the folder universe: naive sample means produce extreme short-saturated weights; the BL prior reproduces $w_{mkt}$ exactly, and a single view tilts it partially.

---

4. Failure Modes & First-Principles Breakdowns

1. “The optimizer will sort out my guesses.” It won’t - it sorts out your errors into extreme bets. The naive-MVO block above is the proof: wildly short one asset you merely guessed too low.
2. Treating the market as efficient. The BL default assumes $w_{mkt}$ is (reverse-)optimal. In an anomaly-laden market the “neutral” prior inherits those anomalies’ biases. This is an assumption, not a theorem - see 05 · Failure Modes.
3. Believing implied returns are “the” expected returns. $\Pi =\delta \Sigma w_{mkt}$ yields the market’s pricing-implied returns under a risk-aversion assumption; they are not a forecast of future realized returns. Confusing the two is the seed of most BL misuse.

---

5. References

• Black & Litterman (1992), Global Portfolio Optimization
• Best & Grauer (1991), On the Sensitivity of Mean–Variance-Efficient Portfolios to Changes in Asset Means, RFS 4(2)
• He & Litterman (1999), Goldman note

---

6. Connected Graph Bridges

• Base: Mean–Variance from Zero · 05 · Estimation-Error Maximizers · Bayes & Priors
• Continue: 02 · Reverse Optimization · Index Hub