Skip to content
Friday, 20 November 2026 · LondonENع
Rayan Azhari.Sustainability · Energy · Carbon · Built Environment · Data Science
Production Quant Trading

Fractional Kelly Position Sizing: Why the Kelly Criterion Is a Leverage, Not a Weight

The Kelly criterion for position sizing does not hand you a 0-to-1 portfolio weight. It hands you a leverage multiple, and for a high-Sharpe, low-vol edge that number comes out at many multiples of your equity. Read it as a weight and you starve a real edge; deploy the raw leverage and you detonate the account. Here is the war-story that taught me to run a fraction of Kelly, and the one rule I now size everything on.

Rayan AzhariChartered Environmentalist, MISEP · 9 min read

There is a units error hiding in the most cited formula in position sizing, and it has cost more retail accounts than any bad signal ever did. You look up the Kelly criterion, you find f* = μ / σ², you compute it for your validated edge, and out comes a number like 9. Every article you have read calls Kelly "the optimal fraction of capital to bet", so you read that 9 as a weight, clamp it to 1.0, and put your whole account on the sleeve feeling conservative. Or you read it literally, deploy nine times your equity, and the first bad week takes the account to zero. Both readings are wrong, and they are wrong in opposite directions, which is exactly why this is worth an essay.

I am Rayan Azhari. I build and run a live, multi-strategy, broker-connected systematic system I call Titan, and I wrote the book these field notes come from. This one is about the single most expensive misconception in sizing, the one none of the top search results state plainly: f* is a leverage, not a portfolio weight. Get that straight and most of the danger evaporates. Miss it and the growth-optimal formula becomes a detonator.

The formula computes a leverage, and it is usually bigger than one

Kelly answers a clean question: what fraction of capital, bet repeatedly, maximises the long-run geometric growth of wealth? For a return stream with per-period mean μ and variance σ², the growth-optimal answer is f* = μ / σ². That is edge over variance. If you met Kelly as a betting formula, f = (pb − q)/b, this is the same object in its continuous, recurring limit.

Now read the units, because that is where the money leaks. f* is not a number between 0 and 1. It is a gross exposure per unit of capital for this strategy in isolation, and for a genuine edge it is routinely greater than 1. The intuition falls out of the denominator. A low-volatility edge, small and steady and consistent, has a tiny σ² on the bottom, so f* blows up. A strategy that earns 12% a year at 4% annual volatility has a Sharpe of 3 and a Kelly leverage of 0.12 / 0.04² = 75. Kelly is telling you, correctly under its own assumptions, to run that edge at seventy-five times gross leverage to compound fastest. That is not a typo and it is not a weight. It is the leverage at which a perfectly known, perfectly Gaussian edge of that shape grows quickest.

This is the correction none of the SERP results say out loud. The good, low-volatility edges, the ones you most want to size well, are precisely the ones whose Kelly number comes out most enormous. So if you have quietly been treating f* as a weight and capping it near 1, you have been massively under-betting your best strategies while feeling prudent. And if you ever take the number at face value and deploy the leverage, the same low-vol edge that produced the huge f* hands you the deepest ruin.

The war-story: a nine that meant nine hundred percent

A candidate sleeve in Titan reported a beautiful, low-volatility out-of-sample return stream. Plugged into the Kelly formula, its f* came out around nine (illustrative of the scale). Full Kelly wanted nine times the account in gross exposure for that one sleeve.

The bug was not in the maths. The maths was right. The bug was in a field name. An early version of my sizing code had named the output full_kelly and then, downstream, treated it as if it were a portfolio weight, a number that belonged in a set that sums to one. So a 9 that was meant to be "nine times leverage, to be scaled way down before deployment" got fed into an allocator that read it as a capital share. A weight of nine is nine hundred percent of the account on a single sleeve. Had that reached the broker unscaled, a normal drawdown on a levered low-vol strategy is the kind of 900% swing that does not leave an account behind to debug.

An internal audit caught it before it touched live capital, and the fix was mostly a rename and a wall between two ideas that had been allowed to blur. The field is now unambiguously a leverage. There is exactly one sanctioned function that converts a leverage into anything that sums to a budget, and it screams the distinction in its own name and docstring. The lesson is mundane and it bites everyone who sizes with Kelly: the number f* is denominated in leverage, a low-vol edge makes it large, and a large leverage read as a weight is a detonator wearing an innocent integer.

Chart

Growth plateaus; drawdown does not

Long-run growth as a function of the Kelly fraction is a flat-topped hill; drawdown rises roughly linearly. Quarter Kelly keeps most of the growth at a quarter of the pain. Illustrative and sanitised.

00.250.50.751.0 (full)1 long-run growth1 drawdown depth
Source: Building a Production Quant Trading System (Titan), illustrative

Why nobody runs full Kelly anyway

Suppose you have the units right and you know your f* of nine is a leverage. You still do not deploy it, for two reasons that compound.

First, full Kelly is growth-optimal only if the edge is known exactly, and it never is. You estimated μ and σ from a finite, noisy sample, and f* is brutally sensitive to that estimate: it scales with μ and inversely with σ². Overestimate the edge by a third, entirely possible after the optimism biases every backtest carries, and you over-lever by a third, into a strategy whose true edge is smaller than you believed. Kelly assumes a certainty you do not have and punishes the shortfall with leverage.

Second, even with the exact edge, full Kelly is unholdable. Its drawdown path is famously vicious: at full Kelly you should expect, in the normal course of things, to lose half your capital at some point, and to spend long stretches deep underwater. The formula optimises the asymptotic growth rate and is completely indifferent to the path taken to get there. But the path is the only thing a human, a clearing firm, and a risk committee actually react to. A growth-optimal strategy that gets switched off at the bottom of its expected fifty-percent trough never lives long enough to realise the growth it was optimal for. Optimal on paper, switched off in practice, is not optimal.

The fix I always use now: a fraction of Kelly

The correction is old, well attested, and the single most useful thing I can hand you: run a fraction of Kelly. Scale the leverage to c · f* with c around a quarter.

Why a quarter buys so much is a happy accident of two different curve shapes, and it is the whole reason this works. Long-run growth as a function of the Kelly fraction is a hill with a flat top: because growth is quadratic near its optimum, backing off from the peak costs you almost nothing at first. Drawdown, by contrast, scales roughly linearly with the fraction: cut the fraction in half and you roughly halve the pain. So the two curves diverge in your favour. Retreating to quarter-Kelly gives up only a sliver of expected growth while cutting the drawdown to roughly a quarter of full Kelly's. You are trading a little bit off the flat top of the growth hill for a large slide down the drawdown line.

Chart

Most of the growth, a quarter of the drawdown

Growth is nearly identical at full and quarter Kelly; the drawdown is not. This chart shows the drawdown depth. Illustrative and sanitised.

Full Kelly50% drawdownQuarter Kelly13% drawdown
Source: Building a Production Quant Trading System (Titan), illustrative

The fraction does double duty. It is not only a drawdown treatment; it is also your margin against the estimation error above. If your edge estimate is inflated by a third, a 0.25 multiplier buys you a large buffer before the over-sizing turns fatal. You are conservative on purpose, twice over: once against a path you could not hold, once against an edge you could not measure exactly. That is why a quarter is a sane default rather than timidity.

The portable rule

If you take one artefact from this piece, take this three-line discipline. Pin it above the sizing code.

  1. f* is a leverage, never a weight. The output of μ / σ² is gross exposure per unit of capital for one sleeve in isolation, and a real low-vol edge makes it much bigger than one. Never feed a raw Kelly number into anything that is supposed to sum to a budget. Convert it deliberately, through one explicit step, and name the two quantities differently so a leverage can never be mistaken for a weight again.
  2. Deploy a fraction of it, not the whole thing. Around a quarter of full Kelly captures the great majority of the long-run growth for a fraction of the drawdown, because growth is flat near the top and drawdown is not. Full Kelly is a ceiling you compute and then step well back from, never a target you deploy.
  3. Size on the pessimistic edge estimate, not the point estimate. f* is a function of a Sharpe, and a Sharpe has error bars. Size on the conservative, haircut reading of your edge, not the flattering centre of its distribution. A sizing decision made on the point estimate is a bet that your luckiest measurement was the truth.

That is the portable core, and it is enough to keep the units error from ever detonating your account. What it deliberately leaves out is the production machinery that turns this principle into a live sizing pipeline: how to gate a sleeve on the lower bound of its leverage so a coin-flip edge earns no capital at all; how the Gaussian μ / σ² lies under fat tails and what empirical number you compute alongside it as a reality check; how vol-targeting re-sizes every bar so the risk you sized for stays constant as markets calm and blow up; the drawdown throttle that halves the effective fraction under stress; and the one explicit, one-directional function that finally converts a set of per-sleeve leverages into portfolio weights that respect a gross budget. Those are the layers between "a fraction of Kelly" and a system you can actually run, and each of them is a place a units error has historically cost real money.

That full sizing chain, the gate, the empirical Kelly, the vol-target overlay, the throttle, and the leverage-to-weight conversion with its thresholds and code, is the subject of the position-sizing chapter in Building a Production Quant Trading System. The book's later half is where the machinery lives, with the schema and the runbook; a sanitised, Apache-2.0 version of the sizing framework is on GitHub if you want to read the shape of it. If you take nothing else from here, take the correction the search results skip: Kelly hands you a leverage. Deploy a fraction of it, on your pessimistic edge, and never let it masquerade as a weight. Read the book for the rest of the chain, and if these field notes are useful, the newsletter is where the next one lands.

This is an engineering essay, not investment advice, and it contains no tradable strategy. All figures are illustrative and sanitised, and the war-stories are about bugs and losses, not profits.

Chart

Growth plateaus; drawdown does not

Long-run growth as a function of the Kelly fraction is a flat-topped hill; drawdown rises roughly linearly. Quarter Kelly keeps most of the growth at a quarter of the pain. Illustrative and sanitised.

00.250.50.751.0 (full)1 long-run growth1 drawdown depth
Source: Building a Production Quant Trading System (Titan), illustrative

Chart

Most of the growth, a quarter of the drawdown

Growth is nearly identical at full and quarter Kelly; the drawdown is not. This chart shows the drawdown depth. Illustrative and sanitised.

Full Kelly50% drawdownQuarter Kelly13% drawdown
Source: Building a Production Quant Trading System (Titan), illustrative

Further reading

Quant field notes, part 19 of 25

Related posts

Purged Cross-Validation and Train/Serve Skew: Four Gates Before an ML Model Trades

When the strategy is a machine-learning model, standard validation lies to you: overlapping labels leak across a naive k-fold split and manufacture a phantom score. Here are the four mechanism-level gates I make any live ML model clear before it touches capital, and the war-story of a model that looked excellent until I purged and embargoed the folds and watched it collapse.

· 9 min

Cannot Trade a US ETF in the UK: Streaming Data, PRIIPs and the Substitute Pattern

Under PRIIPs, a UK or EU retail account can stream a US-listed ETF's market data yet be forbidden from ever holding it, because the fund publishes no Key Information Document. Here is the production pattern I use when my cleanest signal lives on an instrument I am not allowed to trade: compute the signal on the data-only line, trade a UCITS substitute, and match the substitute on currency, not just correlation. The broker catalogue, not the alpha, defines your tradable universe.

· 9 min

Essays in your inbox

New writing on Syria, sustainability and finance, a few times a month.

Unsubscribe anytime. Read by 4,200+ professionals.