5.7.6 Advanced Extensions

Basic Prerequisites: 05 · Failure Modes & Practice and Linear Algebra / Statistics.

---

1. Intuition & Practical Objective

Every page so far sized a single position. Real books hold many assets that co-vary. This page hands you the multi-asset Kelly criterion and shows it is the exact, growth-optimal generalisation that unifies Kelly with the pillar’s mean-variance machinery - multi-asset (continuous) Kelly is the tangency portfolio, scaled by risk aversion. The objective is four ideas:

1. Multi-asset Kelly maximises $\mathbb{E}[\ln W]$ over a weight vector, not a scalar. With covariance $\Sigma$ the growth rate becomes a quadratic form, and the optimum is a closed form in ${\Sigma }^{-1}$.
2. It equals the tangency portfolio. The Kelly-optimal weight vector is $w^{\ast }={\Sigma }^{-1}(\mu -r_{f}1)$, which is exactly the numerator of the MPT tangency portfolio - Kelly is what growth-maximising does to the Markowitz machinery.
3. Uncorrelated assets separate. For independent positions the multi-asset rule reduces to per-asset Kelly $({\mu }_i-r)/s_i^2$; correlation couples them through $\Sigma$.
4. Estimation is the binding constraint. ${\Sigma }^{-1}$ amplifies estimation error (small-eigenvalue blow-up), the exact pathology MPT’s covariance-shrinkage sibling solves - so multi-asset Kelly inherits the need for denoising and fractional de-rating (page 05).

Takeaway. Multi-asset Kelly generalises seamlessly: replace the scalar $f$ with a weight vector $w$, replace $s^2{f}^2$ with $w^T\Sigma w$, and the growth-optimum becomes the tangency portfolio. Growth-optimal investing and mean-variance-optimal investing are the same decision, written in logs instead of quadratics.

---

2. Mathematical Ground Truth & Derivations

2.1 The vector growth rate

Let $w$ be the portfolio weight vector on risky assets, $\mu$ their mean, $\Sigma$ their covariance, and $r_f$ the riskless rate. The continuously-rebalanced growth rate generalises the scalar $g_{\infty }(f)$:

$$g_{\infty }(w)=r_f+w^T(\mu -r_{f}1)-\frac{1}{2}{w}^T\Sigma w.$$

This is a concave quadratic in $w$, so the unique maximiser solves $\nabla g=(\mu -r_{f}1)-\Sigma w=0$:

$$\boxed{w^{\ast }={\Sigma }^{-1}(\mu -r_{f}1)}$$

Bridge to MPT. Normalising $w^{\ast }$ to sum to one recovers the tangency portfolio $w_{\text{tan}}=\frac{{\Sigma }^{-1}(\mu -r_{f}1)}{1^T{\Sigma }^{-1}(\mu -r_{f}1)}$. The Kelly vector $w^{\ast }$ is exactly the tangency portfolio unscaled - i.e. tangency weights × a capital-scaling factor. Growth-optimal == tangent line, and $g_{\infty }(w^{\ast })=\frac{1}{2}(\mu -r_{f}1{)}^T{\Sigma }^{-1}(\mu -r_{f}1)+r_f=\frac{1}{2}{\text{SR}}_{\max }^2+r_f$, the multivariate version of the $S^2/2$ law of page 03.

2.2 Independent assets: the diagonal case

If $\Sigma$ is diagonal (${\sigma }_{ij}=0,i\ne j$), $w_i^{\ast }=({\mu }_i-r_f)/s_i^2$ - each asset independently at its own scalar Kelly fraction. Correlation is the only thing that couples them; two positively-correlated edges should be sized smaller than each alone (less effective diversification), which the $w^T\Sigma w$ cross-terms capture.

2.3 Estimation: the small-eigenvalue amplifier

Because $w^{\ast }={\Sigma }^{-1}(\cdot )$, any noise in $\Sigma$‘s smallest eigenvalues is amplified by the inverse - the same “estimation-error maximiser” pathology as raw MVO (MPT 05). Shrinkage and RMT denoising fix the covariance; the means $\mu$ remain the harder, dominant input error (Chopra–Ziemba).

---

3. Computational Implementation - multi-asset Kelly = scaled tangency

Stdlib plus a tiny Gauss–Jordan inverse (or numpy if available). Two assets, $r_f=0.05$:

Read the last two lines: if the assets were independent, Kelly would put $1.25$ into asset 0 and $1.22$ into asset 1. Because they are positively correlated (${\rho }_{12}=0.012/\sqrt{0.04\cdot 0.09}=0.2$), the coupled optimum shifts weight - less into asset 0 ($0.92$) and less into asset 1 ($1.10$) - exactly the diversification correction ($\Sigma$ cross-terms) that a per-asset shortcut misses. And the normalised $w^{\ast }$ is the tangency portfolio: multi-asset Kelly and the Sharpe-maximising tangency portfolio are the same object.

---

4. Failure Modes & First-Principles Breakdowns

1. ${\Sigma }^{-1}$ amplification of estimation noise. With many assets the smallest eigenvalues of the sample $\Sigma$ are the least trustworthy and the most amplified; raw multi-asset Kelly is a worse estimation-error maximiser than scalar Kelly. Shrink/denoise $\Sigma$ first (Covariance Shrinkage & RMT).
2. Unconstrained leverage. $w^{\ast }$ is not constrained to sum to one and can imply leverage; constraining to long-only or a risk budget fundamentally changes the optimisation and breaks the clean closed form - enter the constraints machinery of MPT.
3. Means dominate and are the hardest input. By the covariance cross-terms, better $\Sigma$ cannot fix a wrong $\mu$; multi-asset Kelly inherits MPT’s “means-are-the-bottleneck” curse (Chopra–Ziemba). Fractional de-rating of the whole vector is mandatory (page 05).
4. Dynamic/regime drift. Parameters estimated over one regime mis-size in the next; multi-asset Kelly assumes a stationary $(\mu ,\Sigma )$. Robust shrinkage + frequent re-estimation + de-rating are the practicable response.

---

5. References

• Thorp, Edward O.: The Kelly Criterion in Blackjack, Sports Betting, and the Stock Market (2006)
• Breiman, Leo: Optimal Gambling Systems for Favorable Games (1961)
• MacLean, Thorp & Ziemba (eds.): The Kelly Capital Growth Investment Criterion (2011)
• Ledoit & Wolf: Improved Estimation of the Covariance Matrix of Stock Returns… (2004)
• Chopra & Ziemba: The Effect of Errors in Means, Variances, Covariances on Optimal Portfolio Choice (JPM 1993)

---

6. Connected Graph Bridges

• Back: 05 · Failure Modes & Practice · Index Hub
• Tangency identity: 03 · Tangency & CAPM · MPT Index
• Covariance fix: Covariance Shrinkage & RMT Denoising · Black–Litterman
• Foundations: 06 · Advanced Extensions (foundations) · Linear Algebra