The Kelly criterion answers a specific question: what fraction of wealth maximizes expected logarithmic growth when a repeatable opportunity has a known probability distribution? A prop challenge asks a different question: can the account reach a target before violating its rules? Confusing those objectives can produce a mathematically correct Kelly fraction that is unusable for the challenge.
Consider a strategy with a genuine 55% probability of winning an amount equal to the amount risked. In the simplest binary model, full Kelly is 10% of current wealth. On a hypothetical $100,000 challenge with a $10,000 loss allowance, that first $10,000 loss can end the attempt. The issue is not an arithmetic error. The sizing formula and the account are solving different optimization problems.
What the binary Kelly formula means
Let p be the probability of winning, q = 1 − p the probability of losing, and b the net profit per dollar risked on a win. A loss removes the entire amount risked. The unconstrained binary Kelly fraction is f* = (bp − q) / b. If shorting or reversing the opportunity is not allowed, a negative fraction means that this long-only betting formulation suggests no allocation, not an instruction to increase risk.
For p = 0.55 and b = 1, the fraction is (0.55 − 0.45) / 1 = 0.10. Winning multiplies wealth by 1 + f, while losing multiplies it by 1 − f. Expected log growth per trade is g(f) = 0.55 × ln(1 + f) + 0.45 × ln(1 − f). Maximizing this expression gives 10% under the stated model.
At 5% risk, expected log growth is about 0.003753 per trade. At 10% it is about 0.005008. At 20% it falls to about −0.000138 even though arithmetic expectancy remains positive. The larger position exposes wealth to a multiplicative penalty from losses. These are theoretical growth rates under known stationary probabilities, not plausible forecasts for a real trading account.
| Policy | Fraction of current wealth | First risk on $100,000 | Share of a $10,000 loss allowance |
|---|---|---|---|
| Quarter Kelly | 2.5% | $2,500 | 25% |
| Half Kelly | 5% | $5,000 | 50% |
| Full Kelly | 10% | $10,000 | 100% |
Quarter Kelly means one quarter of the calculated fraction, not a universal recommendation to risk 2.5%. A different probability or payoff ratio produces a different Kelly estimate. The table also makes clear why fractions that appear moderate relative to a nominal account can be large relative to the contractual loss allowance.
Distinguish wealth compounding from fixed cash risk
Under current-wealth sizing, a 10% loss reduces $100,000 to $90,000 and the next 10% risk is $9,000. Under a fixed 10% of initial balance policy, the next risk remains $10,000. Those processes have different drawdowns and different barrier probabilities. A simulation that fixes the cash amount does not demonstrate the behavior of continuously rebalanced Kelly sizing.
The same distinction applies to ordinary smaller positions. Two consecutive losses at 1% of current equity leave $98,010. Two $1,000 losses leave $98,000. The difference is small at first, but over many trades it changes both upside and downside paths. Record the risk basis explicitly whenever comparing reports.
A further complication is that a prop account's nominal balance may be simulated capital rather than the trader's spendable wealth. The trader's real cash exposure includes fees and future payouts. Treating the displayed balance as personal wealth inside a Kelly objective is a modeling choice, not an economic identity. A contract-level expected-value model can be more relevant to the trader's actual finances.
The probability estimates are rarely known
Kelly is sensitive to the estimated edge. At a 1:1 payoff, an estimated win probability of 55% gives 10% Kelly. At 52%, it gives 4%. At 50%, it gives zero. A few percentage points of estimation error therefore change the allocation substantially. A short or heavily optimized backtest can make the full estimate especially unreliable.
Payoffs also vary. Replacing a distribution containing scratches, tail losses and occasional large wins with one average win and one average loss may distort log growth. For a general return distribution, evaluate the expectation of log(1 + f × return) over admissible outcomes instead of assuming the binary formula is exact. Every modeled outcome must leave positive wealth for the log to exist.
Unobserved losses are another concern. If a strategy's sample contains no major gap, a resampling model cannot invent one automatically. Use economically defensible stress scenarios and acknowledge the limits of the data. A large Kelly estimate based on a truncated loss history is not evidence that the true downside is small.
Translate account constraints into a sizing experiment
For a challenge, define the objective before selecting risk. You might maximize probability of passing within 60 days, maximize expected net payouts over six months or limit the probability of account failure. These objectives can prefer different policies. There is no reason to assume that the risk maximizing one will maximize the others.
Simulate candidate policies against exact target, daily loss and total loss rules. Include minimum trading days, position limits and the treatment of open equity. Stop each path at the first contractually relevant event. Retain unfinished paths. If one policy passes more often only by taking far longer, the horizon and fee structure are part of the interpretation.
Do not optimize a finely spaced risk grid on the same data used to design the strategy and then describe the best point as proven optimal. Use a modest predeclared grid and evaluate the region around the apparent best result. A narrow spike can arise from sampling noise, contract rounding or discrete barrier effects rather than a stable advantage.
A practical cap is a heuristic, not a new Kelly theorem
You can construct a transparent operational cap from several constraints. Suppose the fractional-Kelly allowance is $2,500, remaining total buffer is $10,000, remaining daily buffer is $5,000, and a previously tested fixed-risk ceiling is $1,500. If you budget the total buffer across eight planned full losses and the daily buffer across three, the cash caps are $1,250 and $1,666.67.
The smallest of $2,500, $1,250, $1,666.67 and $1,500 is $1,250. This is a proposed risk budget before contract rounding and an additional reserve for costs or overshoot. It is not a proof of optimality. It combines preferences about how many losses to accommodate with model estimates and program constraints.
After a full $1,250 loss, the total buffer falls to $8,750. Dividing the new buffer by eight yields $1,093.75. The rule reduces risk as room shrinks. It therefore produces a state-dependent policy that must be simulated as such. Testing only its initial $1,250 size does not validate the later behavior.
Be precise about the word survive. If touching the limit is failure, eight identical losses of exactly one eighth of the buffer reach the boundary. To survive all eight without a breach, risk must be strictly smaller, with allowance for costs and execution uncertainty. A formula that uses equality can be a budget reference but is not a safe boundary guarantee.
Handle contracts and simultaneous exposure
Suppose the complete planned loss of one tradable contract, including the cost reserve, is $310. A $1,250 risk cap permits four contracts at $1,240, not 4.03 contracts. If the next cap is $1,093.75, it permits three at $930. Stepwise quantity changes can make the implemented policy differ materially from the smooth fractional curve.
For concurrent positions, allocate the available buffer across the whole portfolio. Four individually acceptable trades can be unacceptable together if they respond to the same event. Adding nominal stop risks is a starting point, but gaps, correlated slippage and cancellation behavior can increase the actual loss. The daily cap belongs to account equity, not to an isolated ticket.
A buffer-sensitive policy may also reach a quantity floor where the smallest position is too large. The correct result is to skip or stop under the defined policy. Silently forcing one contract changes the policy precisely when the account is most constrained. That detail should be visible in both simulation logic and interpretation.
Compare growth and challenge outcomes separately
A useful analysis displays two panels. The first evaluates wealth growth under the selected return model. The second evaluates challenge passing, failure and unfinished probabilities under contractual limits. A policy can perform well on one panel and poorly on the other. That is a genuine tradeoff rather than an inconsistency in the calculations.
Also compare the distribution of cash payouts after passing. Maximizing the chance of reaching a funded stage need not maximize the value of remaining there. Withdrawals may change account headroom, and different funded rules can alter the appropriate risk. Keep the same entries and exits initially so that the experiment isolates sizing rather than mixing two unrelated strategies.
Keep a log of every override to the sizing policy. If actual trading repeatedly exceeds the tested cap, the reported risk distribution no longer describes that behavior. Compare intended and executed risk after each session before interpreting a live drawdown as evidence about the original model.
A disciplined way to use Kelly
- Estimate the complete net outcome distribution from a clearly identified sample.
- Calculate Kelly only for the model and wealth definition it actually describes.
- Test conservative fractions and uncertainty in the estimated edge.
- Apply the actual account constraints in a separate path simulation.
- Round to executable quantities and aggregate simultaneous risk.
- Validate the chosen policy on untouched data and adverse scenarios.
Kelly is useful because it forces the relationship between edge and size into the open. It does not remove estimation error or contractual barriers. For a prop trader, the relevant decision is usually a constrained policy with an explicit horizon and cash objective. Keep Kelly as one analytical reference, then test the policy that can actually be executed.