Kelly Criterion

Introduction

The Kelly Criterion is a position-sizing framework derived from the problem of maximizing the long-run geometric growth rate of capital when the probabilities and payoffs of repeated bets are known or estimated. In trading, Kelly is not a signal generator and it is not a guarantee of maximum account profit. It is a mathematical framework for connecting an estimated edge with the fraction of capital placed at risk.

The most important Expert-level lesson is that Kelly is extremely sensitive to the quality of its inputs. If win probability or payoff assumptions are wrong, the calculated Kelly fraction can be wrong in the most dangerous direction: too large. Kelly therefore belongs inside a broader risk framework that includes Drawdown, Risk of Ruin, Sample Size and uncertainty.

1. The Core Idea

Kelly asks: given a repeatable edge and a defined payoff structure, what fraction of available capital should be exposed to the opportunity if the objective is long-run geometric growth?

The framework comes from repeated favorable opportunities. It is therefore fundamentally different from asking how much to risk on one isolated trade. It assumes that the estimated edge will continue to be available across repeated trials under the model's conditions.

2. The Classic Formula

For a simple binary outcome with decimal-style net odds represented by b, win probability p and loss probability q = 1 − p, the Kelly fraction is:

f* = (bp − q) / b

Here, f* is the theoretically optimal fraction of capital under the simplified model. The formula requires a coherent definition of b and assumes a repeated binary payoff structure.

Example: suppose p = 0.55 and the average net payoff is 1:1, so b = 1. Then q = 0.45 and f* = (1×0.55 − 0.45)/1 = 0.10, or 10%.

3. Why the Formula Is More Dangerous Than It Looks

A 10% result in the example does not mean a professional trader should automatically risk 10% per trade. The formula is the output of a model under specified assumptions. Trading distributions are usually more complex than a binary bet, and the inputs are estimates rather than known constants.

In practice, full Kelly can generate sizing that is operationally aggressive because estimation error is amplified by the formula.

4. Edge Comes First

Kelly cannot manufacture an edge. If the expected edge is zero, the theoretical Kelly fraction is zero under the simple model. If the expected edge is negative, the resulting fraction is negative, which indicates that the modeled opportunity should not be backed in that direction.

This connects Kelly directly to Expectancy and Strategy Edge. Position sizing should follow evidence of edge, not create the appearance of edge.

5. Kelly and Expectancy

Expectancy measures average outcome; Kelly optimizes a specific long-run growth objective under a specified probabilistic model. Positive Expectancy is necessary for a positive Kelly fraction in the simple favorable-bet framework, but the two quantities are not interchangeable.

A larger average edge does not automatically translate into a proportionally larger safe position size. The entire outcome distribution matters, including loss magnitude and probability.

6. Kelly and R-Multiples

R-multiples provide a useful language for separating trade outcome from account scale. If a strategy risks 1R on a defined loss and produces a known distribution of R outcomes, the trader can estimate the edge and then evaluate how different sizing fractions affect compounded equity.

Kelly is most naturally connected to the distribution of returns, not merely to a stop-loss distance. A trade labeled -1R in the journal can still represent different account-level outcomes if the risk fraction changes.

7. Full Kelly and Geometric Growth

The mathematical objective behind Kelly is long-run geometric growth, sometimes described as maximizing expected logarithmic wealth growth. This objective penalizes severe losses because compounding works multiplicatively.

A strategy that maximizes arithmetic average return is not necessarily the strategy or sizing that maximizes geometric growth. Large losses can damage future compounding disproportionately.

8. Why Full Kelly Can Produce Large Drawdowns

Full Kelly is an optimization result under its model, not a drawdown-minimization rule. Even with a genuine edge, the account can experience substantial drawdowns because losing sequences remain possible.

The trader therefore needs to evaluate the Kelly fraction against historical and simulated Drawdown distributions. A mathematically growth-optimal fraction can be psychologically, financially or operationally difficult to sustain.

9. Estimation Error

The most important practical weakness of Kelly is parameter uncertainty. Suppose the true win probability is unknown and the trader estimates it from a finite Sample Size. A small overestimate of p can materially increase the calculated fraction.

The same problem occurs with payoff estimates. If average winners were unusually large in the historical sample, the resulting Kelly fraction may overstate the sustainable edge.

Kelly is therefore only as reliable as the probability and payoff estimates feeding it.

10. Sample Size and Confidence in Kelly Inputs

A Kelly calculation should never be separated from the evidence supporting its inputs. Record the number of trades, the period, market regimes, data quality and how p and payoff were estimated.

A 55% win-rate estimate from a small, highly concentrated sample is not equivalent evidence to a similar estimate supported across a large and diverse sample. This does not mean the larger sample is automatically correct; it means the uncertainty around the estimate must be considered.

11. Kelly and Risk of Ruin

Kelly and Risk of Ruin address different questions. Kelly asks about a growth-optimal fraction under assumptions. Risk of Ruin asks about the probability of crossing a defined capital-loss boundary under assumptions.

They should be used together. If the full Kelly fraction produces drawdowns that are incompatible with the trader's survival threshold, the mathematical optimum is not automatically the operational choice.

This is one of the central professional principles: mathematical optimality and practical survivability are different objectives.

12. Kelly and Drawdown

Position sizing determines how strongly the trade distribution is translated into account-level drawdown. Increasing the Kelly fraction increases the impact of both favorable and unfavorable sequences.

A trader should therefore examine the distribution of maximum drawdowns under candidate Kelly fractions rather than evaluating only the expected growth rate.

13. A Simple Numerical Example

Assume a simplified repeated opportunity with p = 0.60, q = 0.40 and even net payoff b = 1. Full Kelly gives:

f* = (1×0.60 − 0.40)/1 = 0.20, or 20%.

The number is mathematically correct under the simplified assumptions. But if the estimated win probability is actually 55%, the Kelly fraction becomes 10%. If it is 52%, it becomes only 4%. A relatively small estimation change therefore creates a large sizing change.

This sensitivity is why the next lesson, Fractional Kelly, is so important.

14. Kelly Is Not the Same as “Risk Per Trade”

The Kelly fraction describes a capital allocation fraction under a particular model. It should not be confused automatically with the distance between entry and stop or with a broker's margin requirement.

In a trading system, the final position size still depends on instrument price, stop distance, contract/share size, account equity, leverage and the predefined risk model.

A Kelly-derived capital fraction therefore needs to be translated carefully into actual position size.

15. Multiple Strategies and Correlated Bets

Classic Kelly is easiest to understand for one repeated opportunity. A portfolio containing multiple strategies or instruments is more complex because returns can be correlated.

If several trades are effectively the same risk, allocating Kelly fractions independently can produce excessive aggregate exposure. Portfolio-level application therefore requires attention to Correlation Risk, concentration and the joint return distribution.

The central question changes from “How much Kelly for this trade?” to “How much total capital exposure can the portfolio support?”

16. Expert Case Study — The Overestimated Edge

A trader estimates a 60% win rate with 1:1 payoff and calculates a 20% full-Kelly fraction. The estimate comes from 80 historical trades.

After expanding the dataset, the observed win rate is 54%, and the distribution also shows meaningful regime dependence. Under the same simplified 1:1 assumption, the Kelly fraction based on 54% would be 8%.

The original 20% was not a mathematical error. It was an input-estimation problem. The calculation was internally correct for the assumed probability, but the probability estimate was not sufficiently reliable for aggressive sizing.

17. Expert Case Study — Positive Edge, Uncomfortable Path

A strategy has a well-supported positive edge and a theoretical full-Kelly fraction of 12%. Monte Carlo analysis of plausible trade sequences shows that full Kelly can generate deep drawdown paths that would force the trader to abandon the system before the long-run growth objective has time to materialize.

The important observation is that the mathematical model and the operational environment have different constraints. The trader must evaluate the sizing fraction against actual capital tolerance, not only the growth objective.

18. Professional Kelly Workflow

Define the strategy and the repeated opportunity precisely.

Estimate the return distribution from a controlled dataset.

Document win probability and payoff assumptions.

Record Sample Size and regime coverage.

Calculate the theoretical Kelly fraction under the chosen model.

Stress the inputs rather than trusting a single estimate.

Evaluate Drawdown and Risk of Ruin under candidate sizing fractions.

Check aggregate exposure when multiple trades or strategies coexist.

Compare full Kelly with more conservative fractions.

Document the final sizing rule separately from the theoretical Kelly result.

19. Common Mistakes

Treating full Kelly as a universal recommended risk percentage.

Using a small sample to estimate probabilities with false precision.

Ignoring estimation error in win rate or payoff.

Applying a binary Kelly formula to a complex trade distribution without checking assumptions.

Ignoring correlated positions.

Ignoring Drawdown and Risk of Ruin.

Confusing Kelly fraction with stop distance or broker margin.

Optimizing for theoretical growth while ignoring the ability to survive and execute the strategy.

Updating Kelly after every short sequence of wins or losses.

Presenting the calculated fraction without documenting the model that produced it.

20. Expert Exercise — Kelly Sensitivity Map

Choose one strategy and calculate a Kelly fraction under a grid of plausible inputs. For example, vary win probability across a realistic range and vary average payoff separately.

For each combination, record:

Theoretical Kelly fraction.

Expected trade characteristics.

Historical maximum Drawdown.

Simulated adverse drawdown range.

Risk of Ruin against your predefined capital threshold.

The effect of transaction costs and slippage.

The amount of aggregate exposure if several opportunities occur simultaneously.

The objective is not to find the highest number. It is to see how quickly the sizing recommendation changes when the evidence changes.

21. Knowledge Check

1. What objective does the Kelly Criterion optimize in its classic form?

2. What are p, q and b in the simple Kelly formula?

3. Why can a mathematically correct Kelly calculation still be operationally inappropriate?

4. Why is Expectancy related to Kelly but not identical to it?

5. Why does estimation error matter so much?

6. How does Sample Size affect confidence in Kelly inputs?

7. Why should Kelly be evaluated together with Risk of Ruin and Drawdown?

8. Why can independently applying Kelly to correlated positions create excessive exposure?

9. Why must a Kelly fraction be translated carefully into actual position size?

10. What is the purpose of a Kelly sensitivity map?

Copying, reproducing, adapting, distributing, or publicly displaying any part or all of the Academy’s textual content without the author’s prior written consent is prohibited.