The Kelly Criterion estimates the fraction of capital that would maximise long-run expected logarithmic growth under a defined probability and payoff model. In trading, those inputs are estimated rather than known, so the result is best treated as a theoretical sizing benchmark, not an automatic instruction to risk that percentage on the next trade.

Kelly answers Kelly does not answer
What fraction maximises expected log growth under the assumed model? What risk level is safe, comfortable, drawdown-limited or suitable for a specific trader?

What is the Kelly criterion in trading?

The Kelly Criterion is a growth-optimal capital-allocation framework. It was developed for repeated favourable bets where outcome probabilities and payoffs are specified. Its objective is to maximise expected logarithmic wealth growth over repeated opportunities. That is different from maximising average arithmetic return, minimising drawdown or guaranteeing account survival.

For trading, the important qualification is model uncertainty. A win rate, average win and average loss estimated from historical data are not fixed future parameters. Changes in market regime, execution quality, strategy rules or transaction costs can alter the edge and therefore the theoretical Kelly fraction.

Kelly criterion formula and variables

For a simplified binary win/loss model in which a losing outcome loses the defined risked fraction, a common formula is: Edward Thorp’s work on favourable games further developed optimal betting-fraction ideas in repeated favourable settings.

f* = p – q / b = (bp – q) / b

Variable Meaning Trading interpretation Common mistake
f* Estimated Kelly fraction Fraction of bankroll allocated to the modelled risk unit Treating it as broker lot size or notional exposure
p Win probability Estimated probability of a winning outcome Treating historical win rate as known future probability
q Loss probability q = 1 – p Estimating it independently so p + q does not equal 1
b Net payoff ratio Average net win divided by average loss Using gross wins while ignoring costs or inconsistent loss definitions

A negative f* means the stated assumptions do not support a positive Kelly allocation on that side. It should not be silently converted into a small positive trade. Likewise, a result above 100% is a warning to examine model assumptions, leverage and practical constraints, not a routine instruction to risk more than the account value.

How to calculate the Kelly criterion step by step

  1. Estimate the win probability, p, from a sufficiently relevant sample and define exactly what counts as a win and a loss.
  2. Calculate the average net win and average net loss on the same basis. State whether spread, commission, financing and slippage are included.
  3. Divide average win by average loss to obtain the payoff ratio, b.
  4. Calculate q = 1 – p, then apply f* = p – q / b.
  5. Stress-test the result using less favourable p and b values before interpreting the number.

The stress test is not optional in practice. Kelly reacts directly to estimated edge. A modest change in win probability or payoff ratio can produce a much smaller theoretical fraction, or remove the positive allocation entirely.

Worked example: full Kelly, fractional Kelly and sensitivity

The following figures are hypothetical and are included only to show the calculation. They are not typical forex results and are not recommended risk settings.

Input Hypothetical value Calculation / meaning
Win probability, p 55% p = 0.55
Loss probability, q 45% q = 1 – 0.55 = 0.45
Payoff ratio, b 1.5 Average winner = 1.5 x average loser
Full Kelly 25% 0.55 – (0.45 / 1.5) = 0.25
Half Kelly illustration 12.5% 0.25 x 0.5
Quarter Kelly illustration 6.25% 0.25 x 0.25

Under the simplified model, two consecutive full-Kelly losses would leave 56.25% of starting capital, a 43.75% drawdown. Two losses at half Kelly would leave about 76.56%, a 23.44% drawdown. At quarter Kelly, about 87.89% would remain, a 12.11% drawdown. These figures illustrate path sensitivity only; live losses can differ because returns are not perfectly binary and stops can slip or gap.

Sensitivity case p b Raw Kelly f*
Base estimate 55% 1.5 25.00%
Win rate 5 pts lower 50% 1.5 16.67%
Win rate 10 pts lower 45% 1.5 8.33%
Lower win rate 40% 1.5 0.00%

The teaching point is the sensitivity, not the 25% result. If the true win probability is lower than the estimate, or the payoff ratio deteriorates after costs, the theoretical allocation can shrink quickly.

Kelly fraction is not the same as broker position size

The simplified Kelly formula returns a fraction of capital allocated to the defined bet or risk unit. It does not directly return a forex lot size. In a stop-based trading framework, there is an additional translation step:

  1. Account risk amount = account equity x chosen Kelly-based risk fraction.
  2. Determine the planned stop distance in pips, points or price units.
  3. Use the instrument value per pip/point and contract size to convert the risk amount into position units.
  4. Check broker margin requirements, maximum leverage and any position caps.
  5. Remember that gaps and slippage can make realised loss larger than the planned stop-based amount.

Notional exposure can therefore be much larger than the cash amount modelled as risk when leverage is used. Keeping risk fraction, margin and notional exposure separate is essential.

Full Kelly vs fractional Kelly

Approach Fraction applied Growth objective Path risk Main caution
Full Kelly 100% of estimated f* Highest expected log growth under the assumed model Highest of these three examples Highly sensitive to estimation error and drawdown tolerance
Half Kelly 50% of estimated f* Lower expected log growth under same model Lower exposure and usually less severe path variability Not universally optimal or safe
Quarter Kelly 25% of estimated f* Further reduced growth target Lower still under same assumptions May still be too large if the model or tail risk is wrong

Fractional Kelly means scaling the estimated full-Kelly fraction. It is a way to trade some theoretical growth for lower exposure and generally lower path variability under the same model. It is not a repair for a strategy with no edge, and no fraction is automatically suitable for every account or objective.

Why estimation error matters so much

Historical trading data contain sampling error. Even a clean backtest can overstate future win probability or average payoff because the market regime, spreads, slippage, volatility, liquidity and strategy behaviour can change. The Kelly formula amplifies this problem because the output is directly linked to the estimated edge.

  • Small sample: a few unusual wins can inflate p or the payoff ratio.
  • Regime change: a strategy that worked in one volatility or policy regime may behave differently later.
  • Execution drift: spreads, commission, slippage or financing can reduce net payoff.
  • Strategy drift: discretionary changes can make live trades different from the sample used to estimate the inputs.
  • Correlation: several positions that look independent can become one concentrated exposure during stress.

A responsible implementation therefore compares the base estimate with cautious and adverse assumptions instead of treating the historical point estimate as known truth.

Drawdown, variance and practical ruin

Kelly optimises long-run log growth, not the smoothness of the equity curve. Even a correctly specified full-Kelly model can experience deep drawdowns and long recoveries. That does not mean full Kelly must lead to literal bankruptcy, but it can create practical failure long before mathematical wealth reaches zero.

  • Margin close-out can force positions to close before a theoretical recovery path occurs.
  • Gap risk can make a realised loss exceed the modelled loss unit.
  • A drawdown can breach an investment mandate, personal risk limit or broker requirement.
  • A trader may abandon a strategy after losses that are mathematically possible but psychologically or financially intolerable.

Research on risk-constrained Kelly methods exists precisely because classical growth optimisation does not itself impose a maximum-drawdown constraint. A separate risk constraint changes the problem being solved.

Trading costs, slippage and variable outcomes

Simple model assumption Live-market complication Why it matters
Fixed win/loss sizes Returns vary across trades One average payoff ratio can hide tail losses or rare outsized wins
Known p and b Inputs are estimated and may change Theoretical f* can be materially wrong
Clean entry/exit Spread, commission and slippage Net edge can be lower than the backtest estimate
Loss capped at planned amount Gaps and stop slippage Real loss can exceed the modelled unit
Independent opportunities Correlated positions Portfolio risk can be understated by separate single-trade Kelly sizing

How to use a Kelly calculator responsibly

A useful Kelly calculator should make assumptions visible instead of hiding them behind a single percentage. At minimum, it should accept win probability, average win and average loss, then show the raw fraction and sensitivity to worse inputs.

  • Flag negative Kelly rather than converting it to a positive position.
  • Flag outputs above 100% and explain why leverage/model assumptions need review.
  • Warn when the sample size is small or payoff estimates are unstable.
  • Allow a fractional-Kelly multiplier, but label it as user-selected rather than optimal.
  • Keep lot-size conversion separate unless account equity, stop distance, value per pip/point and broker constraints are provided.

When Kelly is a poor fit

Kelly is most informative when the probability/payoff model is reasonably stable and the objective genuinely resembles long-run growth. Direct use is a poor fit when the edge cannot be estimated credibly, when drawdown limits dominate the objective, or when outcomes contain tail risks the simplified model does not capture.

Situation Role for Kelly Main reason
Stable, well-sampled strategy Use as a theoretical benchmark Inputs may be informative enough for sensitivity analysis
Uncertain or changing strategy Use cautiously, mainly for scenario analysis p and b may not be stable
Strict capital-preservation or drawdown mandate Do not use raw Kelly directly Objective differs from unconstrained log-growth maximisation

Fixed-risk, volatility-targeted or explicitly drawdown-constrained frameworks can be easier to govern when model uncertainty is high. They solve different problems, so they should be compared by objective rather than treated as universal winners.

UK leveraged-trading context

For UK retail contracts for difference and rolling spot forex, broker margin rules and FCA retail protections still matter regardless of any Kelly calculation. As checked on 29 September 2026, FCA rules for retail CFDs include leverage limits by underlying, a 50% margin close-out rule and negative balance protection. Those protections do not make an aggressive Kelly-derived position appropriate, and a theoretical risk fraction should never override the actual legal-entity, product and margin terms of the account.

Kelly criterion checklist for traders

  • Define the model first: what counts as a win, loss and risk unit?
  • Use net outcomes after the costs included in the strategy analysis.
  • Calculate p, q and b transparently, then apply the correct formula for the simplified model.
  • Stress-test lower p, lower b and higher costs.
  • Separate the Kelly risk fraction from notional exposure and broker lot size.
  • Check drawdown tolerance, gaps, leverage, correlation and margin constraints.
  • If using fractional Kelly, state the fraction and why it was chosen.
  • Re-estimate inputs when the strategy or market regime materially changes.
Live signals on Telegram

Turn Trading Calculations Into Real Trading Decisions

Use your calculated risk levels, position sizing, and technical analysis together with live forex trading signals and market commentary from The Forex Complex.

325,000+ traders inside the community
Join The Forex Complex Learn how it works
Wesley, Founder of TFXC - The Forex Complex
Wesley · Founder of TFXC

Frequently asked questions

What does the Kelly Criterion optimise?

It maximises expected logarithmic growth under the assumed probability and payoff distribution. That is different from minimising drawdown, maximising arithmetic return or guaranteeing account survival.

What is the formula for Kelly position sizing?

In the simple binary model, f* = p – q/b, where p is win probability, q = 1 – p and b is the average net win divided by the average loss. Real trading with variable outcomes may require a more general expected-log model.

Can I use my historical win rate in the Kelly formula?

You can use it as an estimate, but it is uncertain and may not persist. Use a relevant sample, include costs where appropriate and test lower win-rate and payoff assumptions.

Why can full Kelly be so large?

If the model estimates a strong edge, full Kelly can be large because it targets long-run log growth rather than low drawdown. A large output is a reason to inspect assumptions and practical constraints, not proof that large risk is suitable.

Is half Kelly safer than full Kelly?

It uses less exposure and generally reduces path variability under the same model, but suitability still depends on return distribution, leverage, gaps, correlation and input accuracy.

Can Kelly prevent risk of ruin?

Not automatically. Classical Kelly does not impose a maximum-drawdown constraint. Practical ruin can still occur through margin close-out, gap losses, model error or an intolerable drawdown.

What if the Kelly result is negative?

Under the stated inputs, the model does not support a positive allocation on that side. Recheck the assumptions and costs rather than forcing a positive trade.

Can a Kelly calculator give me forex lot size?

Only if it also has the inputs needed to translate account risk into units, such as equity, stop distance and pip or tick value. The raw Kelly fraction alone is not a lot-size calculation.