Multi-Asset Kelly & Ruin

The last two lessons sized a single bet: one favorable wager, one fraction $f^* = \mu/\sigma^2$, and the sobering discovery that betting more than full Kelly trades growth for variance until, eventually, it trades growth for nothing. But you almost never hold a single position. A real book is a basket — a dozen stocks, three strategies, a sleeve of bonds — all live at once, all sloshing around together. The question stops being “what fraction?” and becomes “what vector of fractions?” — and the answer pulls in everything you already know about covariance, correlation, and diversification.

This is the capstone. We take Kelly from one bet to many, watch correlation quietly rewrite the optimal sizes, count the true cost of leverage, and then nail down the most important guardrail in the whole subject: the mathematics of ruin. Past a certain point, more aggression doesn’t just add risk — it guarantees you end up with nothing, edge or no edge.

Many bets at once: from a number to a vector

Analogy. Sizing one bet is like setting the volume on a single speaker: one knob, one sweet spot. Sizing a portfolio is like mixing a whole band — drums, bass, guitar, vocals. You can’t tune each channel in isolation, because they bleed into each other: crank two instruments that always play the same riff and you’ve just made one loud, muddy track, not a richer mix. The optimal board is a set of levels that accounts for how every channel interacts.

Definition. With $n$ assets you no longer choose a scalar fraction $f$ — you choose a vector of fractions $\mathbf{f} = (f_1, f_2, \dots, f_n)$, where $f_i$ is the fraction of your bankroll allocated to asset $i$. Each asset has an excess return (its return above the risk-free rate), collected into a vector $\boldsymbol{\mu}$, and the assets’ co-movements live in the covariance matrix $\Sigma$ — the same $n \times n$ matrix from portfolio theory, with variances $\sigma_i^2$ on the diagonal and covariances $\sigma_{ij} = \rho_{ij}\sigma_i\sigma_j$ off it.

The single-bet formula was $f^* = \mu/\sigma^2$ — edge divided by variance. The multi-asset formula is the exact same shape, just promoted to linear algebra: divide the edge vector by the variance matrix. “Dividing by a matrix” means multiplying by its inverse, $\Sigma^{-1}$. That one substitution carries the whole story, because $\Sigma^{-1}$ is where correlation does its work.

Pitfall. A tempting shortcut is to size each asset with its own scalar Kelly, $f_i = \mu_i/\sigma_i^2$, and call it a day. That ignores every off-diagonal term in $\Sigma$ — i.e. all the correlations — and systematically over-bets whenever your assets move together (which, in practice, they almost always do). The whole point of the vector formula is that it refuses to treat your positions as independent islands.

When it matters

The leap from scalar to vector matters the instant you hold more than one position whose returns are not perfectly independent — which is to say, always. A long-only equity book, a market-neutral pair, a multi-strategy fund: every one of them needs the matrix, not a stack of separate scalars.

The vector Kelly formula: f* = Σ⁻¹μ

Analogy. Think of $\Sigma^{-1}$ as a de-correlating lens. Raw, your assets overlap and double-count their shared risk. Passing the edge vector $\boldsymbol{\mu}$ through $\Sigma^{-1}$ untangles the overlap: it credits each asset for the unique edge it brings after stripping out what its correlated neighbors already cover, and discounts redundant exposure. The output is the set of stakes that, deployed simultaneously, maximizes long-run growth.

Definition. For a vector of excess returns $\boldsymbol{\mu}$ and covariance matrix $\Sigma$, the growth-optimal Kelly weights are

This is the continuous (log-normal) form of multi-asset Kelly, and it is one of the most quietly profound results in the field — because it is exactly the direction of the tangency portfolio (the maximum-Sharpe portfolio) from mean-variance theory. Unconstrained mean-variance optimization, with leverage allowed, points you at the very same mix of assets. Kelly’s contribution is to tell you how far to lean into that direction: it picks the tangency portfolio as the right shape, then scales it to the size that maximizes geometric growth.

So Kelly and Markowitz aren’t rivals — they’re describing the same vector. Markowitz finds the best direction in risk-return space; Kelly fixes the best magnitude along it. The formula $\mathbf{f}^* = \Sigma^{-1}\boldsymbol{\mu}$ rolls both into one line.

A worked two-asset example: how correlation reshapes sizing

Numbers make the magic concrete. Take two symmetric assets, each with excess return $\mu = 8\% = 0.08$ and volatility $\sigma = 20\% = 0.20$, so each has variance $\sigma^2 = 0.04$. We’ll size them under two correlation regimes and watch the weights move.

The clean shortcut. For two identical assets (same $\mu$, same $\sigma$), symmetry forces $f_1 = f_2 = f$, and the vector formula collapses to a single tidy expression:

The $(1+\rho)$ in the denominator is the entire correlation story in one term. Crank $\rho$ up and you divide by a bigger number, so each stake shrinks; push $\rho$ down toward $-1$ and the denominator collapses, so each stake balloons.

Case 1 — uncorrelated ($\rho = 0$). Two genuinely independent edges:

Each asset gets a weight of $2.0$, for a gross exposure of $4.0$ — i.e. 4× leverage. With no overlap, Kelly happily piles into both independent edges.

Case 2 — highly correlated ($\rho = +0.8$). The same two edges, but now they move almost in lockstep:

Each weight falls to about $1.11$, gross exposure about $2.22$ — barely more than half the uncorrelated case. High correlation shrank the optimal size by nearly half, even though each asset’s standalone edge and volatility never changed. Why? Because two bets that win and lose together aren’t two bets — they’re closer to one doubled-up bet, with no diversification to dilute the joint swings. Kelly sees the missing diversification and pulls in its horns.

Case 3 — negatively correlated ($\rho = -0.5$), for contrast.

Negative correlation is a hedge baked into the basket: when one zigs, the other tends to zag, smoothing the combined ride. That extra stability lets Kelly size up dramatically — to $4.0$ each here, double the uncorrelated stake.

Misconception. “More winning positions always means I should deploy more capital.” Not so. Adding a correlated clone of a bet you already hold barely diversifies you, so it adds risk faster than it adds edge — and Kelly’s optimal total exposure can actually fall as you tack on near-identical positions. It’s the independence of edges, not the count of them, that earns you the right to size up.

When it matters

Correlation-aware sizing is the difference between a risk team that survives a crisis and one that doesn’t. In calm markets, correlations look modest and books look diversified; in a panic, correlations lurch toward $+1$ and that “diversified” book behaves like one giant position. Sizing with the true (often stress-elevated) $\Sigma$ is what keeps the joint drawdown survivable.

What leverage really costs

Analogy. Leverage is a turbocharger: it multiplies whatever the engine is already doing — including blowing it up. The Kelly formula, in its bare form, assumes you can bolt on as much turbo as the math wants, for free, with no downside until the very end. Reality bolts on a fuel bill and a kill-switch that can trip at the worst possible moment.

Definition. Total Kelly leverage is the gross exposure $\sum_i |f_i|$. Notice the worked example already blew past $1.0$: gross exposure of $4.0$ means borrowing three dollars-worth for every dollar of your own. The textbook formula $\mathbf{f}^* = \Sigma^{-1}\boldsymbol{\mu}$ treats that borrowing as frictionless — infinite, costless, never recalled. Three real costs break that fantasy:

1. Borrowing isn’t free. You pay a financing rate on borrowed capital. The honest fix is to feed the formula returns measured in excess of your borrowing cost, not just in excess of the risk-free rate. A 4× position whose financing rate eats most of the edge is nothing like a 4× position funded for free — the net $\boldsymbol{\mu}$ that actually drives $\mathbf{f}^*$ is smaller than the gross one.
2. Margin calls force the worst-timed sales. The single-period formula sees only start and end; it is blind to the path. But a leveraged book that draws down can trip a margin call mid-path and be liquidated at the bottom, locking in a loss the math assumed you’d ride through. Leverage converts a temporary paper drawdown into a permanent, forced realization.
3. Liquidity and slippage. Unwinding a large levered position in a stressed market moves the price against you. The frictionless formula assumes you trade at the screen; reality charges a toll that scales with size.

Misconception. “Kelly told me 4× is optimal, so 4× is safe.” The formula optimizes expected log growth under its own assumptions — frictionless leverage, no forced exit, perfectly known $\boldsymbol{\mu}$ and $\Sigma$. Strip those assumptions and the leverage Kelly hands you is an upper bound on what’s defensible, not a target. Practitioners run a fraction of it precisely because the costs above bite hardest exactly when you’re already hurting.

When it matters

The leverage caveat dominates any strategy that needs borrowing to reach its Kelly size — leveraged ETFs, futures-overlay funds, basis and carry trades. The more leverage the math demands, the more violently these frictions punish a mistake, which is exactly why prime brokers and risk committees cap gross exposure far below the theoretical optimum.

Over-betting and the cliff, made rigorous

Analogy. Imagine a dial from 0 to 2-and-beyond, where 1 is full Kelly. As you turn it up past 1 you go faster — for a while. But the road curves back on itself: at the “2” mark you arrive exactly where you started (zero growth), and past it you’re driving in reverse, accelerating toward the cliff. The terrifying part is how symmetric and innocent the early over-betting feels — right up until it isn’t.

Definition. Bet a fraction $k$ of full Kelly (so $k = 1$ is full Kelly, $k = 0.5$ is half-Kelly, $k = 2$ is double). The long-run growth rate obeys the clean quadratic

Read off the landmarks. At $k = 1$: $2 - 1 = 1$, full growth — the peak. At $k = 0$: zero growth (you didn’t bet). At $k = 2$: $2(2) - 2^2 = 4 - 4 = 0$, growth is exactly zero. And for $k > 2$: the term $2k - k^2$ goes negative, so your bankroll’s long-run growth rate is negative — you are guaranteed to be ground down to ruin, no matter how large your edge. Edge cannot save you once you’re past twice Kelly; the geometry of compounding does the killing.

The symmetry, spelled out. The parabola $2k - k^2$ is symmetric about its peak at $k = 1$. So under-betting at fraction $k$ and over-betting at fraction $(2 - k)$ deliver the identical growth rate. Half-Kelly ($k = 0.5$) and 1.5×-Kelly ($k = 1.5$) both yield $2k - k^2 = 0.75$ of peak growth — same expected log growth on paper.

But — and this is the entire moral of the lesson — those two are not equivalent in risk. The over-bet (1.5×) achieves that growth with far more volatility, far deeper drawdowns, and far more exposure to estimation error and ruin. Same growth, wildly more pain. So if you must err, err on the low side: you give up the same growth as the over-bettor, but you keep your sanity, your drawdowns, and your capital. The rule writes itself: when in doubt, bet less.

Risk of ruin and drawdown: why full Kelly still hurts

Analogy. Full Kelly is a Formula 1 car: it posts the fastest possible lap time over a long race, but it spends that race millimeters from the wall, and a single twitch sends it into the barrier. Staying on the optimal line and surviving it are two different skills — and most drivers, sensibly, lift off the throttle.

Definition. Even at or below full Kelly, the path is brutal. Full Kelly maximizes growth but accepts enormous interim drawdowns as the price. A well-known rule of thumb: under full Kelly, the probability that your bankroll ever falls to some fraction $a$ of its starting value is roughly $a$ itself.

Worked intuition. Set $a = 0.5$. The rule says there’s about a 50% chance your bankroll halves at some point under full Kelly — a coin-flip that you’ll watch half your money evaporate before the compounding pays off. Set $a = 0.1$ and there’s roughly a 10% chance of a 90% drawdown along the way. These aren’t tail freak events; they’re baseline behavior of the growth-optimal strategy. Most humans (and most fund mandates) cannot stomach a 50% drawdown, which is the practical reason almost nobody runs full Kelly.

Now stack on estimation error. Your $\boldsymbol{\mu}$ and $\Sigma$ are estimates, riddled with noise. If you’ve overestimated your edge — and edges are systematically overestimated, because the strategies you choose to run are the ones that looked good — then the $f^*$ you computed is already an over-bet relative to the true optimum. You think you’re at $k = 1$; you’re really at $k = 1.4$, eating the over-bettor’s drawdowns for none of the imagined growth. Fractional Kelly (betting $k = 0.25$ to $0.5$ of the formula) is the practical safeguard: it sacrifices a sliver of theoretical growth (recall $k = 0.5 \Rightarrow 0.75$ of peak) to buy a huge reduction in drawdown and a fat margin of safety against having mis-estimated your edge in the first place.

The chart drives it home — three bettors, the same sequence of outcomes, sized at half-Kelly, full-Kelly, and an over-bet. Watch the half-Kelly bettor climb almost as high as full-Kelly with a fraction of the white-knuckle dips, while the over-bettor rockets up and then a perfectly ordinary losing streak detonates the account.

Pitfalls: where multi-asset Kelly bites back

The vector formula is elegant on paper and treacherous in production. Three traps account for most blowups.

(a) Σ and μ are estimated — and Σ⁻¹ amplifies the noise. This is the deepest trap. Both inputs are sampled from finite, noisy history. Inverting $\Sigma$ then magnifies that noise: $\Sigma^{-1}$ is dominated by the matrix’s smallest eigenvalues, which correspond to the near-redundant, highly-correlated directions your data measures worst. Tiny estimation errors there get blown up into enormous, unstable, wildly over-levered weights — the identical “error-maximization” pathology that makes raw mean-variance optimization notorious. The cures are the same: shrink the estimates toward something stable, constrain the weights (caps, no-leverage limits), or simply run fractional Kelly, which scales the whole noisy vector down and tames the damage.

(b) Kelly is a long-run result. The growth-optimality of Kelly is an asymptotic statement — it wins as the number of bets goes to infinity. Over a short horizon, or when you have a fixed withdrawal need (a pension paying out, a fund facing redemptions), full Kelly is far too aggressive: a deep early drawdown you’d eventually recover from in the limit can be fatal if you must draw cash through it. Finite horizons and forced withdrawals both argue for sizing well below the formula.

(c) Ruin is path-dependent — the average lies. A strategy can have a perfectly healthy average (mean) outcome while most individual paths get wiped out, because a handful of astronomically lucky paths drag the mean up. The arithmetic mean is blind to ruin; the median path, and the fraction of paths that survive, are what you actually live. This is why Kelly optimizes the log (geometric) growth rate, not the arithmetic mean — and why you should judge a sizing rule by its survival distribution, not its headline expected return.

When it matters

These pitfalls are why real risk teams cap position sizes well below the theoretical Kelly number — typically a quarter to a half, with hard leverage and concentration limits stapled on top. Multi-strategy funds run a budget of fractional-Kelly sleeves rather than one all-in vector. The thread running through it all — that survival is path-dependent and the mean can hide a graveyard — is exactly what the risk-of-ruin and portfolio-optimization topics develop next: turning “don’t get wiped out” from a slogan into a measured, managed constraint.

Putting it together

Multi-asset Kelly is the single-bet rule grown up. Sizing many positions at once means choosing a vector $\mathbf{f}^* = \Sigma^{-1}\boldsymbol{\mu}$ — edge vector divided by the covariance matrix — which points in the exact direction of the tangency / maximum-Sharpe portfolio, scaled for geometric growth. Correlation is the hidden hand: positively correlated bets are redundant, so Kelly shrinks them; negative correlation is a built-in hedge that lets you size up. The bare formula assumes frictionless leverage, but real borrowing costs, margin calls, and slippage all argue for running below it. And the iron law of ruin is non-negotiable: growth scales as $(2k - k^2)\,g(f^*)$, hitting zero at twice Kelly and going negative beyond — over-bet and your edge cannot save you. Because your $\boldsymbol{\mu}$ and $\Sigma$ are noisy estimates that $\Sigma^{-1}$ amplifies into over-levered nonsense, the practitioner’s answer is always the same: bet a fraction of Kelly, and when in doubt, bet less.

That closes the Kelly & Geometric Growth ladder: from why we compound in log-space, through sizing a single edge, fractional-Kelly prudence, and now the full multi-asset vector and the hard wall of ruin. The throughline never changed — grow geometrically, survive first, and when the math and your nerves disagree, bet less. From here the platform turns to risk of ruin and portfolio optimization proper, where “don’t get wiped out” graduates from a maxim into a measured, enforced constraint on every position you hold.