Kelly sizing: the fragile bridge from edge to exposure

Optimize the right quantity

Kelly sizing maximizes expected logarithmic wealth in a specified model. It is not a method for discovering whether an edge exists, and it does not maximize the chance of making money next month. Logarithms penalize outcomes that leave little wealth, capturing the compounding consequence of losses. The optimization still depends completely on the assumed probabilities, payoffs and feasible positions.

For a simple repeated wager that wins b units per unit staked with probability p and loses the entire stake otherwise, the unconstrained fraction is f = p − (1 − p)/b. This formula is not directly interchangeable with a stop-loss percentage or a leveraged instrument's margin requirement. Its loss definition is part of the model, not an administrative setting on a trading platform.

The binary formula follows from the wealth multipliers, not from the win rate alone. Risking fraction f produces wealth of 1 + bf after a win and 1 − f after a loss, relative to starting wealth. Expected log growth is therefore p ln(1 + bf) + (1 − p) ln(1 − f). Differentiating and setting the slope to zero gives the stated optimum. This derivation also exposes its limits: wealth must stay positive in every modeled outcome, and any borrowing or shorting constraint changes which fractions are feasible.

Small probability errors create large sizing errors

Consider even-money outcomes and an estimated win probability of 55%. The formula gives a 10% fraction. If the true probability is 51%, the model's optimum is only 2%; if it is below 50%, a positive position has no modeled edge. A few percentage points of estimation error can therefore transform the apparent optimum into substantial overexposure.

An observed win rate is not a known probability. One hundred trades from overlapping signals can contain much less than one hundred independent observations. The payoff distribution may also change after fees, missed fills and unusual losses. Before considering any sizing result, describe how the estimate was obtained, its plausible uncertainty range and which adverse outcomes are absent from the sample.

Growth does not mean a tolerable journey

A strategy can have positive expected log growth while suffering deep interim drawdowns. Long-run claims do not guarantee survival under a margin call, a withdrawal requirement or a psychological limit. An investor with an imminent cash need does not have the unlimited horizon assumed by many simplified demonstrations.

Fractional Kelly scales the calculated position downward, often to reduce sensitivity and drawdown, but the fraction is not a universal safety certificate. Half of a seriously overstated optimum can still be too large. Research on risk-constrained Kelly formulations makes drawdown restrictions explicit rather than assuming that a growth objective automatically represents all preferences. Those models also retain assumptions that need independent scrutiny.

Multiple opportunities share one balance sheet

Applying a separate full Kelly calculation to each signal can silently reuse the same capital. Simultaneous positions may lose together, and their joint payoff distribution determines the portfolio problem. Three trades linked to the same currency shock are not three independent repetitions merely because the symbols differ.

Practical constraints include borrowing costs, minimum contract sizes, concentration limits and losses beyond the planned exit. A binary win-or-lose model suppresses these details. Build a scenario table for total account wealth, including correlated gaps and funding needs, before translating a theoretical fraction into units. Cash reserved for expenses or emergencies should not be casually treated as risk capital merely to make a larger position look proportionate.

Use the formula as a sensitivity exercise

A useful educational application is to calculate a range of fractions across plausible probabilities and payoffs, then inspect how unstable the answer becomes. If a modest deterioration removes the edge, the important conclusion is fragility, not the largest number in the table. Negative modeled growth is also informative: decreasing exposure cannot manufacture an economic advantage that is absent after costs.

Keep research uncertainty separate from personal capacity for loss. A model can be statistically promising yet unsuitable for a particular horizon or funding arrangement. Record the objective, assumptions and binding constraints together. The purpose of the exercise is to reveal what must be believed for a position to be justified, not to produce an authoritative percentage that bypasses judgment.

Check the loss model before debating the fraction

A quoted margin requirement is not the maximum amount at risk. If an instrument needs $1,000 of initial margin but can lose $3,000 through a gap before it is closed, treating the trade as a wager that loses only $1,000 changes the problem. The same mistake occurs when a stop distance is treated as an enforceable payoff boundary. The fraction in a growth calculation must refer to the economic loss distribution actually being assumed, not the cash a platform initially asks to hold.

Even the simple binary model makes the cost of a losing sequence clear. At a fixed 10% fraction of current wealth, ten consecutive losses leave 0.9¹⁰ ≈ 34.9% of starting wealth: a drawdown of roughly 65.1%. Returning from that level to the starting amount requires about 186.8% growth. At a 2% fraction, the same sequence leaves about 81.7%, requiring approximately 22.4% growth to recover. These calculations describe specified paths, not the probability that either path will occur.

The result also depends on whether size is reset after each outcome. Risking 10% of current wealth shrinks the dollar stake after losses. Risking a fixed $100 from an initial $1,000 account is a different process; after five losses, that unchanged stake is 20% of remaining wealth. A backtest that computes a fraction once and then keeps the dollar position fixed has not simulated repeated fractional sizing. Minimum trade sizes can prevent idealized fractional reductions as an account becomes smaller.

Rare outcomes deserve explicit scenarios. In a log-growth model, an outcome that wipes out all wealth has log wealth of negative infinity; omitting it because it was absent from a short history can radically change the apparent optimum. This is not a reason to assign invented probabilities with spurious precision. It is a reason to distinguish estimated everyday outcomes from unmodeled failure modes, and to ask whether the proposed exposure survives the latter at all.