The Kelly Criterion: Why 1% Risk Is Mathematically Defensible

The most common way competent traders end their careers is not a faulty thesis. It is a correct thesis sized too aggressively. They identify real edges, catch legitimate breakouts, and ride genuine trends, and then they give it all back in a single quarter because the size of their average position was mathematically incompatible with the variance of their own system. This failure mode is not psychological and it is not unusual. It is the predictable output of a specific, well-understood equation.

That equation was published in 1956 by a Bell Labs physicist who was not thinking about trading at all. John L. Kelly Jr. was trying to work out how much a gambler with inside information should wager on a sequence of bets in order to maximize the long-run growth of his bankroll. The answer turned out to be neither intuitive nor conservative. It was the mathematical boundary between compounding and ruin — a sharp cliff, not a gentle slope, beyond which even a genuinely profitable strategy destroys capital faster than it creates it.

Kelly's result has since been generalized, debated, refined, and half-applied by generations of investors. What follows is the derivation, the translation into trading terms, and the specific chain of real-world adjustments that takes a theoretically optimal Kelly fraction of 20% or 30% down to the 1% risk-per-trade number that disciplined swing traders actually use. That number is not a timid rule of thumb. It is what survives when estimation error, correlation, and the asymmetric pain of drawdown are all priced into the formula honestly.

The Problem Kelly Was Actually Solving

Imagine a repeated bet with known probabilities. Each round you can stake any fraction of your current bankroll. If you win, you are paid some multiple of what you staked. If you lose, you forfeit the stake. Your edge is positive but finite, and you will play this game thousands of times. The question is not whether to bet, but how much to bet.

The naive answer — bet whatever maximizes expected dollar return on the next round — is catastrophically wrong, and the reason it is wrong is the single most important idea in position sizing. Expected dollar return is an arithmetic mean across parallel universes. Your bankroll, however, does not live in parallel universes. It lives in one sequence, multiplied one round at a time. The right objective is not to maximize the average outcome but to maximize the geometric growth rate of a single path through time.

This distinction sounds academic until you run the numbers. A coin-flip that pays 3-to-1 on heads and takes your whole stake on tails has a positive expected value at any bet size up to 100% of bankroll — the arithmetic mean is always positive. And yet if you stake 100% of bankroll every round, your probability of still having money after twenty rounds is roughly one in a million. Expected value is up; you are broke. The arithmetic mean is correct about each round in isolation and completely useless as a guide to compounding.

Kelly's insight was to ask the right question. He did not ask which bet size maximizes the next payoff. He asked which bet size maximizes the exponent of long-run wealth growth, and he showed that this exponent has a unique interior maximum. Bet too little and you leave growth on the table. Bet too much and the variance of ruin overwhelms the mean of profit. Somewhere in between sits a single fraction that is, in a precise mathematical sense, the best possible long-run sizing — the one that maximizes the expected logarithm of terminal wealth.

Deriving Kelly from First Principles

Consider the simplest case. You have a bet that wins with probability p, paying b units of profit per unit staked, and loses with probability q = 1 − p, forfeiting the stake entirely. You stake a fraction f of your current bankroll W every round. After a win, your bankroll becomes W(1 + bf). After a loss, it becomes W(1 − f).

After N rounds containing approximately pN wins and qN losses, your bankroll is, to first order:

Taking the natural logarithm and dividing by N gives the per-round geometric growth rate G(f). This is the exponent that governs compounding, and it is the quantity Kelly chose to maximize:

Setting the derivative of G(f) with respect to f equal to zero and solving produces one of the most compact and consequential results in applied mathematics:

The formula reads almost too cleanly to feel serious, but it is the whole game. The probability of winning, minus the probability of losing divided by the payoff ratio, gives the fraction of bankroll to stake. If the right-hand side is negative or zero, there is no edge to exploit and the optimal fraction is zero — do not bet. Anything positive tells you the exact size that maximizes geometric compounding, given the inputs you believe.

The Trader's Translation

For a trader, Kelly's abstract “stake” maps onto a very specific quantity: the dollar amount risked per trade, which is the distance between entry and stop multiplied by position size. On a winning trade, that risk unit returns some multiple b — the average R-multiple on wins, where R is the initial risk. On a losing trade, the full risk unit is forfeited when the stop is hit. The mapping to Kelly's bet is exact, provided stops are honored and slippage is modeled separately.

That means the Kelly fraction f* translates directly into percent of account risked per trade. A Kelly fraction of 0.20 says: risk 20% of equity on every single trade. A Kelly fraction of 0.05 says: risk 5% on every trade. This is the clean, lossless translation — and it is also where most traders first encounter how violent Kelly's prescription really is.

The two inputs a trader needs are p, the win rate of the system, and b, the ratio of average winner to average loser measured in R. The table below shows the full Kelly fraction for realistic combinations. Cells marked with a dash indicate a negative-expectancy system — no edge, no bet.

The table is a cold bath. A well-documented momentum system running at a 45% win rate with a 2.5R payoff — an entirely plausible set of statistics for a disciplined VCP-style trader — is told by full Kelly to risk 23% of equity per trade. At 55% and 3R, the prescription climbs to 40%. Almost no retail trader would seriously consider sizing at these levels, and almost no retail trader should. The reason is not that Kelly is wrong. The reason is that full Kelly answers a different question than the one traders actually need to answer.

The Over-Betting Cliff

Kelly is a maximum, which means the growth rate is locally flat near f* — slightly under-betting barely costs you anything. But the function is profoundly asymmetric. Over-betting is punished at an accelerating rate, and at exactly twice the Kelly fraction something remarkable happens: the geometric growth rate returns to zero. Beyond that point, a system with positive expected value becomes negatively compounding in the long run. The arithmetic mean says you are winning. The bankroll disagrees.

The shape of the curve explains why full Kelly is dangerous even when its inputs are perfectly known. The peak is flat on one side and a cliff on the other. Under-betting by 20% costs you perhaps 4% of maximum growth. Over-betting by 20% costs you more than 30% of maximum growth. Over-betting by 50% costs you almost everything. And crucially, because the downside grows quadratically with error while the upside grows linearly, any uncertainty in p or b — and in trading there is always uncertainty — pushes the rational operating point meaningfully to the left of the theoretical peak.

The Case for Half-Kelly

The conventional response to Kelly's asymmetry is to divide the recommended fraction by two. Half-Kelly is not an arbitrary haircut — it is the result of three distinct arguments that happen to converge on the same answer.

The growth-rate argument. At exactly half-Kelly, the long-run geometric growth rate is 75% of its maximum. That sacrifices only one quarter of theoretical compounding. In exchange, the variance of returns drops by roughly half, because the variance of a fractional-Kelly strategy scales with f² while its mean growth scales with f(2 − f/f*). Cutting the bet size in half sacrifices a small amount of mean growth for a large reduction in volatility — a favorable trade for anyone who cares about the distribution of outcomes, not just its expected value.

The drawdown argument. Drawdowns under full Kelly are legendary in their severity. A full-Kelly strategy will, with meaningful probability, experience a 50% drawdown somewhere along its compounding path — this is a feature of the mathematics, not a failure of the system. Half-Kelly reduces the probability of a 50% drawdown by more than an order of magnitude and softens the worst observed drawdowns substantially. For a human operating a system under stress, this difference is not cosmetic. It is the difference between continuing to execute the plan and abandoning it mid-drawdown.

The estimation-error argument. This is the argument that matters most, and it is the one most traders undervalue. The Kelly formula assumes that p and b are known with certainty. In trading, they are not known — they are estimated from a finite sample, and both estimates are noisy. A widely-cited result in the Kelly literature shows that if the true win rate is unknown and you estimate it with standard statistical error, the optimal fraction to bet is materially less than the point estimate of Kelly applied to your best guess. In the limit, with enough uncertainty about the true parameters, the rational Kelly fraction collapses toward half of the naive estimate — sometimes less.

All three arguments push in the same direction. Half-Kelly is not a concession to nerves. It is what Kelly's own mathematics recommends once you acknowledge that you do not know your own edge with certainty.

From Half-Kelly to 1%

Half-Kelly is still not the number most disciplined swing traders actually use. Professional allocators, institutional risk desks, and experienced discretionary traders almost universally converge on a working figure closer to 0.5% to 1% of capital risked per trade. This number is roughly one-quarter of half-Kelly for a strong system and one-eighth for an exceptional one. Each step of that additional deflation has a specific justification.

Parameter non-stationarity. The Kelly formula assumes that p and b are constants. In markets they are not. Volatility regimes change, correlation structures shift, the character of trend leadership rotates, and the statistical edge of any given system oscillates around its long-run mean. A system that delivers a 45% win rate and 2.5R over a decade will deliver materially different numbers in any given year. Sizing at half-Kelly on the decade-average assumes the edge is stable; the reality is that you must size for the worst plausible regime you will actually trade through.

Correlation across open positions. Kelly's formula applies to a single independent bet. A swing trader running five positions simultaneously does not have five independent bets — those positions are typically drawn from the same market regime, often the same few leading sectors, and they rise and fall together during broad-market corrections. Effective Kelly exposure for a portfolio of correlated positions is substantially lower than the sum of the individual Kellys. A rule of thumb widely used by institutional traders is to divide single-position Kelly by the effective number of correlated positions open at any time, which for concentrated momentum portfolios is often far below the nominal count.

The pain-of-drawdown asymmetry. A 50% drawdown requires a 100% recovery to reach breakeven. A 75% drawdown requires a 300% recovery. The mathematics of loss recovery is not symmetric, and neither is the psychological experience of trading through it. Kelly's objective function — maximize expected log wealth — treats drawdowns and recoveries as analytically equivalent, but a human operator does not. Accepting a smaller long-run growth rate in exchange for shallower drawdowns is not inefficient; it is a rational utility adjustment that Kelly's purely mathematical objective does not capture.

Execution and model friction. Real trades incur slippage, partial fills, and commissions that are not in the theoretical expected-value calculation. Stops are sometimes violated intraday and filled worse than the theoretical stop price. A 1R loss in the model is occasionally a 1.3R loss in the account. The formula assumes clean arithmetic that reality routinely violates.

The cumulative effect of these adjustments can be summarized in a simple deflation chain. Consider a realistic momentum system:

The 1% figure falls out the bottom of this chain. It is not a superstitious round number and it is not an arbitrary beginner's rule. It is the result of taking full Kelly and applying each of the corrections that honest estimation requires. A trader operating at 1% risk on a system with these statistics is running at roughly one-twentieth of theoretical Kelly — and that conservatism is not excessive. It is the price of admission for compounding capital across regimes whose character cannot be known in advance.

When 1% Is Still Too Much

There are conditions under which even 1% is aggressive. Traders in the first hundred trades of a new system have no statistically meaningful estimate of their own win rate or payoff and should size a fraction of what a proven system would justify — 0.25% to 0.5% is a reasonable bootstrap range. Traders returning from a drawdown are carrying a damaged psychological edge in addition to any statistical one, and should size down until a new equity high is made. Traders in confirmed market corrections face degraded breakout quality and elevated failure rates — the edge itself shrinks, and sizing should follow.

Conversely, there are conditions under which sizing above 1% is defensible: exceptional setups with unusually tight stops that compress the dollar loss on failure, highly liquid names with minimal slippage assumption, and systems with extensive, stationary track records that justify a higher quarter-Kelly floor. These conditions are rare and the adjustment should be modest. Doubling from 1% to 2% is a meaningful change; tripling it is the sort of decision that asks whether one is still operating within Kelly's safe zone at all.

The Broader Principle

Kelly's contribution to trading thought is not a formula. It is a reorientation. Before Kelly, the implicit framework for sizing was some combination of personal risk tolerance, anecdotal rules from older traders, and the arithmetic mean of expected returns. After Kelly, sizing is a mathematical problem with a mathematical boundary, and the question becomes not “how much can I afford to risk” but “how much does the geometry of compounding permit.”

The shift matters because it changes what a drawdown means. A trader who blows through the Kelly boundary is not unlucky — they have been running a negative compounding strategy disguised as a winning one, and the drawdown is merely the mathematics catching up with the sizing. A trader who stays well inside the boundary can absorb long losing streaks without existential damage, because the geometric growth rate remains positive even when the short-term arithmetic is punishing.

Every serious sizing framework — from institutional risk parity, to the volatility targeting used by CTA funds, to the 1% rule taught in every disciplined trading course — is ultimately a descendant of Kelly's insight that compounding is governed by geometry, not arithmetic. The specific numbers differ. The logic does not.