5.3.2 Reverse Optimization & the Implied Equilibrium Returns

Basic Prerequisites: Tangency Portfolio & CAPM and Linear Algebra.

---

1. Intuition & Practical Objective

This page derives the one number that makes Black–Litterman work: the implied equilibrium return vector $\Pi$. Instead of estimating expected returns with noisy statistics (the hopeless job that breaks MVO), you reverse-engineer them from a fact you can actually observe: the market capitalization weights $w_{mkt}$. The move is pure bookkeeping - invert Markowitz on the market portfolio and ask which expected returns would have produced those weights as the optimizer’s answer.

The objective in one line: “If I believe the market-cap portfolio is held because it’s optimal, then the expected returns that justify it are $\Pi =\delta \Sigma w_{mkt}$ - and those, not my noisy sample means, are the only sane starting prior.” This converts an unobservable quantity (true expected returns) into a deterministic function of observables $\Sigma$ and $w_{mkt}$ plus one parameter $\delta$ you can set or estimate.

Three intuitions:

1. Weights are the data; returns are the output. In Markowitz you feed returns and get weights. Reverse optimization feeds observed market weights and gets the returns that rationalize them. The market tells you what it prices in - not necessarily what will happen, but what a mean-variance agent must believe to hold the market.
2. It is the CAPM’s equilibrium in disguise. In CAPM equilibrium the market portfolio is tangency; the implied returns are exactly the CAPM-consistent expected returns. So the BL prior is a market-neutral, CAPM-grounded prior - the clean textbook starting point before you add opinion.
3. Calibrate $\delta$ from the market, don’t guess it. $\delta$ is the risk-aversion that sizes both the prior and the final portfolios. The classic trick: back it out from the market’s own Sharpe, $\delta =\frac{{\mu }_{mkt}-r_f}{{\sigma }_{mkt}^2}$.

---

2. Mathematical Ground Truth & Derivations

Forward Markowitz. Maximizing risk-adjusted expected return

$$U(w)=w^T\mu -\frac{\delta }{2}{w}^T\Sigma w,$$

the first-order condition is ${\nabla }_{w}U=\mu -\delta \Sigma w=0$, i.e. the optimal weight vector satisfies

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

Reverse optimization. We observe $w_{mkt}$ (market-cap weights) but not $\mu$. Postulate that an investor with risk-aversion $\delta$ does hold $w_{mkt}$ as optimal, so $w^{\ast }=w_{mkt}$. Invert the first-order condition:

$$\underset{\text{implied returns}}{\underbrace{\Pi }}=\delta \Sigma w_{mkt}.$$

The defining identity is that re-solving forward with $\Pi$ recovers the market exactly:

$$\frac{1}{\delta }{\Sigma }^{-1}\Pi =w_{mkt}.$$

Why $r_f$ sits inside $\Pi$ (and never re-enters the weights). Write the FOC with excess returns, $\mu -r_{f}1=\delta \Sigma w_{mkt}$: the implied returns are $\Pi =r_{f}1+\delta \Sigma w_{mkt}$, i.e. $\Pi$ is a total-return vector whose excess part $\Pi -r_{f}1=\delta \Sigma w_{mkt}$ is exactly what reverse-optimizes. When you re-solve forward you must therefore use the excess-return tangency form

$$w_{\text{tan}}=\frac{{\Sigma }^{-1}(\Pi -r_{f}1)}{1^T{\Sigma }^{-1}(\Pi -r_{f}1)}=\frac{{\Sigma }^{-1}(\delta \Sigma w_{mkt})}{1^T{\Sigma }^{-1}(\delta \Sigma w_{mkt})}=\frac{w_{mkt}}{1^T{w}_{mkt}}=w_{mkt}.$$

The $\delta$ and $r_f$ cancel; the equilibrium is recovered exactly. The folklore “drop $r_f$” is precise in this sense: $r_f$ lives in $\Pi$ but drops out of the weights.

Calibrating $\delta$. Multiply the identity by $w_{mkt}^T$: $w_{mkt}^T\Pi =\delta w_{mkt}^T\Sigma w_{mkt}$. The LHS is the market’s excess expected return ${\mu }_{mkt}$, the last term the market variance ${\sigma }_{mkt}^2$, giving

$$\delta =\frac{{\mu }_{mkt}}{{\sigma }_{mkt}^2}({\mu }_{mkt}=w_{mkt}^T\Pi \text{ is already the excess return, since }\Pi =\delta \Sigma w_{mkt}\text{ is the excess vector}).$$

Requiring the model to price the market at its own observed Sharpe pins $\delta$ down, removing the entirely free dial.

The prior. Returns are modeled as random with mean $\Pi$ and covariance $\tau \Sigma$:

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

The $\tau$ (typically $\approx 0.02$–$0.10$) scales the estimate uncertainty of the prior mean far below the assets’ own variances - the prior is confident about where returns sit on average, not about their realized dispersion.

---

3. Computational Implementation - reverse the optimizer

Computes $\Pi$, verifies the reverse-optimization identity round-trips to $w_{mkt}$, and calibrates $\delta$ two ways (given vs. market-implied).

---

4. Failure Modes & First-Principles Breakdowns

1. $\delta$ is not free - it rescales both $\Pi$ and the final $w^{\ast }$. Fix $\delta$ by requiring the model to price the market at its own observed Sharpe (here $\delta ={\mu }_{mkt}^{\text{excess}}/{\sigma }_{mkt}^2=2.5$, recovering the given value) - a poorly chosen $\delta$ rescales every implied return; see 06 · Advanced Extensions for the consequence.
2. $w_{mkt}$ must be investable and observable. If you use a proxy index with poor coverage, or weights that include illiquid positions, the “equilibrium” you invert is garbage. Reverse optimization is only as clean as its inputs.
3. Equilibrium is an assumption. The identity $\frac{1}{\delta }{\Sigma }^{-1}\Pi =w_{mkt}$ is constructed to hold, not derived from data. If the market is not efficient, $\Pi$ is the return a model with wrong beliefs would price, not the true opportunity set.

---

5. References

• Black & Litterman (1992)
• He & Litterman (1999)
• Idzorek (2005)
• Connection: Sharpe (1964), Capital Asset Prices

---

6. Connected Graph Bridges

• Base: Tangency & CAPM · Efficient Frontier · Linear Algebra
• Continue: 03 · The BL Formula · Index Hub