Position Sizing Is Risk Management of The Mathematics of the Kelly Criterion

A good investment idea can still become a bad portfolio decision if it is sized incorrectly.

That is the central problem behind position sizing.

An investment process must answer two separate questions. First, does the opportunity have a positive expected edge? Second, given that edge and the uncertainty surrounding it, how much capital should be allocated?

The Kelly Criterion provides one rigorous answer to the second question. Its objective is to maximize the expected logarithmic growth of wealth over repeated opportunities. The framework originates from J. L. Kelly Jr.'s 1956 paper, A New Interpretation of Information Rate.

But Kelly is often presented too simply. It is not a universal formula for “the right position size,” nor is it a guarantee of superior investment performance. It is an optimization framework whose result depends on the assumptions used to describe returns, probabilities, payoffs and, in a portfolio setting, dependence between assets.

That distinction is where mathematics meets risk management.

The Position-Sizing Problem

Consider an investor who allocates a fraction (f) of capital to an investment opportunity with return (R).

The wealth process can be written as:

$$W_{t+1}=W_t(1+fR_{t+1})$$

After multiple periods:

$$W_T=W_0\prod_{t=1}^{T}(1+fR_t)$$

The important feature is that wealth compounds multiplicatively.

This means that the arithmetic average of periodic returns does not fully describe the long-run behavior of capital.

Taking logarithms:

$$\log W_T = \log W_0+ \sum_{t=1}^{T}\log(1+fR_t)$$

The multiplicative wealth process has now become additive in log space.

That transformation leads directly to the Kelly objective.

Kelly as a Log-Growth Optimization Problem

Define expected logarithmic growth as:

$$g(f)=E[\log(1+fR)]$$

The Kelly fraction is the value of (f) that maximizes this function:

$$\boxed{ f^*=\arg\max_f E[\log(1+fR)] }$$

This is the fundamental Kelly equation.

The objective is therefore not simply to maximize expected return.

It is to maximize expected logarithmic wealth growth under the specified return distribution.

Edward O. Thorp's work on portfolio choice and the Kelly Criterion formalized this capital-growth interpretation and discussed its relationship with Markowitz mean-variance methods. Thorp also emphasized that maximizing expected log wealth does not automatically maximize every possible terminal-wealth utility function.

That distinction is important because “optimal” always depends on the objective being optimized.

The First-Order Condition

For an interior solution, differentiate the growth function with respect to (f):

$$\frac{dg(f)}{df} = E\left[ \frac{R}{1+fR} \right]$$

At the optimum:

$$\boxed{ E\left[ \frac{R}{1+f^*R} \right]=0 }$$

This is the general first-order condition for the one-period Kelly problem.

The allocation must also respect the domain:

$$1+fR>0$$

for the outcomes considered by the model.

This condition matters because a mathematical optimization cannot be separated from the possibility of capital destruction. A position-sizing model that ignores the lower bound of wealth is incomplete.

The Classic Binary Kelly Formula

The familiar Kelly formula is a special case of the general framework.

Suppose an investment has:

probability of winning (p);

probability of losing (q=1-p);

net payoff (b) for every unit of capital risked on a win.

The expected logarithmic growth is:

$$g(f) = p\log(1+bf)+q\log(1-f)$$

Differentiating:

$$\frac{dg}{df} = \frac{pb}{1+bf} - \frac{q}{1-f}$$

Setting the derivative equal to zero gives:

$$\boxed{ f^*=\frac{bp-q}{b} }$$

or:

$$\boxed{ f^*=p-\frac{q}{b} }$$

This is the standard binary Kelly equation.

The formula contains three pieces of information:

$$p$$

the probability of success,

$$q$$

the probability of failure, and

$$b$$

the net payoff relative to the amount risked.

The size of the position therefore depends not only on whether an investment has an edge, but on the magnitude and structure of that edge.

A Simple Example

Consider a hypothetical strategy with:

$$p=55\%$$

and:

$$q=45\%$$

Assume an even-money payoff:

$$b=1$$

The Kelly fraction becomes:

$$f^* = 0.55-\frac{0.45}{1}$$

Therefore:

$$\boxed{f^*=10\%}$$

Under the assumptions of this simplified binary model, the theoretical growth-optimal allocation is 10% of capital.

But the number should not be interpreted as a recommendation to invest 10%.

It is the output of a model.

That distinction is fundamental.

If the 55% probability estimate is wrong, the calculated Kelly fraction is also wrong.

Kelly Is Not a Maximum-Return Formula

The phrase “Kelly maximizes returns” is convenient but mathematically imprecise.

Kelly maximizes:

$$E[\log(1+fR)]$$

under the specified model.

It does not necessarily maximize:

arithmetic expected return;

single-period return;

minimum volatility;

minimum drawdown;

Sharpe ratio;

terminal wealth under every utility function;

or any investor's preferred risk-adjusted objective.

This is one reason Kelly belongs within a broader portfolio-construction framework rather than being treated as a standalone trading rule.

Why Compounding Matters

Consider a hypothetical sequence of returns:

$$+50\%,\quad -50\%$$

The arithmetic average is:

$$\frac{50\%-50\%}{2}=0\%$$

But suppose initial capital is

After the first period:

$$100(1.50)=150$$

After the second:

$$150(0.50)=75$$

The investor finishes with

The average return was zero, yet capital declined by 25%.

This is the fundamental reason geometric compounding matters.

Kelly's logarithmic objective captures that compounding process directly.

Kelly and Portfolio Theory

Kelly should not be viewed simply as an alternative to mean-variance portfolio theory.

The two frameworks use different optimization objectives.

A mean-variance portfolio can be represented using:

$$E[R_p]=w^\top\mu$$

and:

$$\sigma_p^2=w^\top\Sigma w$$

where:

(w) is the vector of portfolio weights;

(\mu) is the vector of expected returns;

(\Sigma) is the covariance matrix.

The generalized Kelly problem is instead:

$$\boxed{ w^* = \arg\max_w E[\log(1+w^\top R)] }$$

The distinction is therefore not simply “Kelly versus Markowitz.”

The more precise distinction is:

What objective should determine the portfolio allocation?

Modern portfolio construction also requires investors to consider objectives, constraints, risk tolerance, diversification and the interaction between assets. CFA Institute's 2026 Level III curriculum explicitly places asset allocation, portfolio construction and derivatives and risk management within the core portfolio-management framework.

The (\mu/\sigma^2) Approximation

A frequently cited approximation to Kelly sizing is:

$$f^*\approx\frac{\mu}{\sigma^2}$$

The relationship comes from a second-order approximation of the logarithm.

For sufficiently small (x):

$$\log(1+x) \approx x-\frac{x^2}{2}$$

Therefore:

$$E[\log(1+fR)] \approx fE[R] - \frac{1}{2}f^2E[R^2]$$

Let:

$$\mu=E[R]$$

Then:

$$g(f) \approx f\mu-\frac12f^2E[R^2]$$

Differentiating:

$$\frac{dg}{df} \approx \mu-fE[R^2]$$

The approximate optimum is:

$$f^* \approx \frac{\mu}{E[R^2]}$$

Since:

$$E[R^2]=\sigma^2+\mu^2$$

and under assumptions where the mean is small relative to the variance term:

$$E[R^2]\approx\sigma^2$$

we obtain:

$$\boxed{ f^*\approx\frac{\mu}{\sigma^2} }$$

The qualification is important.

Kelly is not generally equal to expected return divided by variance.

That relationship is an approximation under specific conditions.

The Real Risk: Estimation Error

The theoretical equation is elegant.

The inputs are not.

In a real investment process, the true probability (p), expected return (\mu), volatility (\sigma), and covariance matrix (\Sigma) are unknown.

They have to be estimated.

Instead of:

$$p$$

the investment model works with:

$$\hat p$$

Instead of:

$$\mu$$

it uses:

$$\hat\mu$$

And instead of:

$$\Sigma$$

it uses:

$$\hat\Sigma$$

The distinction between the true parameter and its estimate is a source of model risk.

A Small Change in the Edge Can Produce a Large Change in Position Size

Return to the binary example.

Suppose:

$$b=1$$

If:

$$p=55\%$$

then:

$$f^*=10\%$$

Now suppose the true probability is only:

$$p=52\%$$

Then:

$$q=48\%$$

and:

$$f^* = 0.52-0.48 = 4\%$$

The estimated probability changed by only three percentage points.

The theoretical Kelly allocation changed from:

$$10\%$$

to:

$$4\%$$

That is a 60% reduction in the calculated position size.

This illustrates a central problem with Kelly sizing:

precision in the formula does not imply precision in the underlying knowledge.

A spreadsheet can calculate 9.83% to two decimal places.

That does not mean the model knows the optimal position size to two decimal places.

Fractional Kelly

One practical response to uncertainty is fractional Kelly.

Instead of allocating the full theoretical Kelly fraction:

$$f^*$$

the investor uses:

$$\boxed{ f_{\text{fractional}}=cf^* }$$

where:

$$0<c<1$$

For a theoretical Kelly fraction of 10%:

These are illustrations, not universal recommendations.

There is no mathematical rule that says 50% Kelly is always appropriate.

The appropriate fraction depends on the reliability of the estimated edge, the investment horizon, portfolio constraints, liquidity, leverage and the investor's objective.

Fractional Kelly should therefore be understood as a deliberate reduction in exposure to the model's estimated edge—not as a magical risk-control parameter.

Kelly in a Multi-Asset Portfolio

The single-position formula becomes insufficient when several positions interact.

For a portfolio return vector (R) and portfolio-weight vector (w):

$$W_{t+1}=W_t(1+w^\top R)$$

The portfolio Kelly problem becomes:

$$\boxed{ w^* = \arg\max_w E[\log(1+w^\top R)] }$$

Now dependence between assets becomes central.

Two investments may each appear attractive on a standalone basis while creating excessive portfolio risk when combined.

If:

$$\rho_{12}\approx0$$

their joint behavior may be materially different from a portfolio where:

$$\rho_{12}\approx1$$

Expected return has not necessarily changed.

The risk interaction has.

This is why position sizing cannot be separated from portfolio construction.

CFA Institute's portfolio-management curriculum similarly emphasizes that risk and return characteristics, correlations and investor objectives all matter when constructing portfolios.

Correlation Is Only Part of the Problem

Correlation is useful, but it does not fully describe financial dependence.

Financial return distributions can exhibit:

fat tails;

volatility clustering;

changing dependence;

non-normality;

regime changes;

extreme observations.

These characteristics matter because a Kelly allocation is only optimal relative to the distribution assumed by the model.

A model that understates the probability of extreme losses can produce a position that appears optimal under normal conditions but becomes excessive during stress.

That is why professional risk management extends beyond a single expected-return or volatility estimate.

Transaction Costs Matter

The theoretical Kelly calculation can produce a gross allocation.

Implementation produces a net result.

Conceptually:

$$R_{\text{net}} = R_{\text{gross}}-C$$

where (C) may include:

commissions;

bid-ask spreads;

market impact;

slippage;

financing costs;

taxes where applicable.

If a strategy's expected edge is small, even modest costs can materially change the economics.

This becomes especially important when the position-sizing model rebalances frequently.

A mathematically optimal allocation that cannot be executed at the assumed price is not an operationally optimal allocation.

Leverage Changes the Risk Profile

Kelly can generate allocations that require leverage.

Leverage introduces another set of constraints:

margin requirements;

financing costs;

liquidity requirements;

forced deleveraging;

gap risk;

counterparty exposure.

Therefore:

$$\text{Theoretical Kelly} \neq \text{Permissible Portfolio Exposure}$$

A portfolio manager may impose leverage or concentration limits even when an unconstrained optimization suggests a larger position.

This is consistent with the broader risk-management structure of the FRM curriculum, which includes quantitative analysis, valuation and risk models, market-risk measurement and management, liquidity and treasury risk, and investment management.

Kelly Needs a Risk-Control Layer

A practical quantitative process can be represented as:

Investment Edge

↓

Kelly Optimization

↓

Parameter Uncertainty

↓

Portfolio Risk

↓

Risk Constraints

↓

Execution Constraints

↓

Monitoring

The optimization is only one layer.

A portfolio may also operate under limits for:

volatility;

Expected Shortfall;

drawdown;

concentration;

leverage;

liquidity;

turnover.

The objective is not to replace Kelly with arbitrary limits.

The objective is to prevent an optimization result from being interpreted outside the assumptions under which it was calculated.

The AM CAPITAL Position-Sizing Framework

A disciplined position-sizing process can be organized into nine steps.

1. Identify the Edge

Estimate the expected return, probability or payoff distribution associated with the strategy.

1. Estimate the Distribution

Determine the assumptions governing:

$$\mu,\quad\sigma,\quad\Sigma$$

or the appropriate probability and payoff parameters.

1. Calculate Theoretical Kelly

Solve:

$$f^* = \arg\max_f E[\log(1+fR)]$$

or the corresponding portfolio optimization.

1. Quantify Uncertainty

Ask how reliable the inputs are.

A point estimate without an assessment of uncertainty is incomplete.

1. Apply Fractional Scaling

Where justified:

$$f_{\text{scaled}}=cf^*$$

1. Apply Portfolio Constraints

Evaluate:

concentration;

leverage;

liquidity;

volatility;

tail risk;

mandate restrictions.

1. Include Implementation Costs

Estimate:

$$\text{spread} + \text{slippage} + \text{market impact} + \text{financing}$$

1. Stress Test

Recalculate the allocation under adverse assumptions.

For example:

$$p\downarrow$$ $$\sigma\uparrow$$ $$\rho\uparrow$$ $$C\uparrow$$

1. Validate Out of Sample

The sizing process should ultimately be evaluated using information that was not used to estimate its parameters.

This final step is essential.

Backtesting Kelly Without Fooling Yourself

A common research error is to estimate Kelly parameters using an entire historical dataset and then evaluate the resulting strategy on the same data.

That can make the strategy appear more robust than it actually is.

A more defensible framework is:

Parameter estimation

↓

Kelly calculation

↓

Risk constraints

↓

Forward period

↓

Performance measurement

↓

Re-estimation

This is the basic logic behind walk-forward analysis.

At each point in time, only information that would actually have been available at that point should be used.

The same principle applies to:

expected returns;

probabilities;

volatility;

correlations;

transaction costs;

portfolio constraints.

What Should Be Measured?

Terminal wealth alone is not enough to evaluate a position-sizing framework.

A quantitative evaluation should consider multiple dimensions.

The correct set of metrics depends on the investment objective and portfolio mandate.

No single statistic captures every dimension of risk.

What Happens When the Model Is Wrong?

The most useful Kelly analysis may not be the base-case calculation.

It may be the failure analysis.

If the edge is overestimated:

$$\hat p>p$$

the calculated position may be too large.

If volatility is underestimated:

$$\hat\sigma<\sigma$$

realized risk may exceed modeled risk.

If correlation is underestimated:

$$\hat\rho<\rho$$

positions that appear diversified may become concentrated during stress.

If transaction costs are underestimated:

$$\hat C<C$$

net returns may fall below the model's assumptions.

If the regime changes:

$$P_{\text{future}}(R) \neq P_{\text{historical}}(R)$$

the historical model may no longer describe the investment environment.

The formula can therefore be mathematically correct while the investment decision is economically wrong.

That is the difference between model correctness and model validity.

Full Kelly Versus Fractional Kelly

Full Kelly Fractional Kelly

Allocation (f^) (cf^)

Exposure to estimated edge Full Reduced

Sensitivity to model assumptions Higher Lower exposure

Growth objective Full Kelly objective Reduced exposure to that objective

Model risk Remains Remains

Need for validation High High

Fractional Kelly is therefore not a substitute for research.

Reducing a position does not make an inaccurate forecast accurate.

It simply changes how much capital is exposed to that forecast.

What Kelly Does Not Solve

Kelly does not determine whether the forecast itself is correct.

It does not solve:

Forecasting risk

A poor estimate of future returns remains a poor estimate.

Data risk

Bad or biased data produce bad model inputs.

Regime risk

Historical relationships may change.

Liquidity risk

A theoretical position may not be executable at the modeled price.

Transaction-cost risk

Costs may be higher than expected.

Correlation risk

Dependence can change during periods of market stress.

Governance risk

Portfolio mandates may impose restrictions that an unconstrained optimizer does not consider.

Tail risk

The result depends on how the return distribution is specified.

These are not peripheral concerns.

They determine whether the mathematical result can be used responsibly.

Position Sizing as a Risk-Management Decision

The broader investment process can therefore be summarized as:

$$\boxed{ \text{Position Size} = f( \text{Edge}, \text{Uncertainty}, \text{Portfolio Risk}, \text{Liquidity}, \text{Costs}, \text{Constraints} ) }$$

Kelly provides the mathematical framework for one part of that equation:

$$\text{Expected Logarithmic Growth}$$

Risk management determines whether the resulting exposure remains appropriate once uncertainty, diversification, liquidity, leverage and implementation constraints are considered.

This distinction is consistent with the structure of modern professional investment education. CFA Institute's current Level III framework places asset allocation, portfolio construction, performance measurement and derivatives/risk management among its core areas, while GARP's FRM curriculum combines quantitative analysis and valuation/risk models with market-risk measurement and management.

The practical conclusion is straightforward.

Position sizing is not an administrative step after investment selection. It is the mechanism through which an estimated investment edge becomes actual portfolio exposure.

Conclusion

The Kelly Criterion is powerful because it turns a qualitative question—how much capital should be committed to an edge?—into a mathematical optimization problem.

Its central expression is:

$$\boxed{ f^*=\arg\max_f E[\log(1+fR)] }$$

For a simple binary payoff structure, that reduces to:

$$\boxed{ f^*=\frac{bp-q}{b} }$$

But the simplicity of the formula should not obscure the complexity of its inputs.

The probability may be estimated incorrectly.

The payoff distribution may change.

Volatility may be understated.

Correlations may rise during stress.

Transaction costs may be higher than expected.

Liquidity may disappear when it is needed most.

And the future distribution of returns may simply differ from the historical distribution used to build the model.

That is why the most important question is not:

“What is the Kelly fraction?”

It is:

“How much confidence should a portfolio place in the assumptions that produced the Kelly fraction?”

For quantitative investors, that is where position sizing becomes risk management.

References

Kelly, J. L. Jr. (1956). A New Interpretation of Information Rate. Bell System Technical Journal, 35(4), 917–926.

Thorp, E. O. (1975). Portfolio Choice and the Kelly Criterion. In Stochastic Optimization Models in Finance, pp. 599–619. Academic Press.

CFA Institute. CFA Program Level III — 2026 Curriculum. Asset Allocation, Portfolio Construction, Performance Measurement, and Derivatives and Risk Management.

GARP. 2026 FRM Study Materials and Curriculum. Quantitative Analysis, Valuation and Risk Models, and Market Risk Measurement and Management.

Google Search Central. A new resource for optimizing for generative AI in Google Search. Google, May 15,

1. Google emphasizes valuable, unique, non-commodity content and states that foundational SEO practices remain relevant to generative AI features in Search.