Why full Kelly is too aggressive for trading
The Kelly criterion has an odd reputation: mathematically optimal, and widely discounted in practice. That is not timidity. Full Kelly is optimal only if you know your edge exactly, and a trader’s edge is an estimate — usually from a sample too small to pin it down. Everything wrong with full Kelly in practice follows from that one gap.
This piece is for readers who have run the Kelly criterion calculator, seen a surprisingly large number, and want to know why people discount it. It does not tell you what fraction to use.
All figures below are worked examples, computed from the formula
f = (b·p − q) ÷ b, where b is the reward/risk ratio, p the win rate and
q = 1 − p.
Your win rate is a range, not a number
Kelly takes p as given. You do not have p. You have a count of wins over a
count of trades, and that ratio carries sampling error that shrinks only with
the square root of the sample size.
For a true win rate around 45%, the standard error of the observed rate is:
| Trades in sample | Standard error | Rough 95% range |
|---|---|---|
| 30 | ±9.1 pts | ±17.8 pts |
| 50 | ±7.0 pts | ±13.8 pts |
| 100 | ±5.0 pts | ±9.8 pts |
| 200 | ±3.5 pts | ±6.9 pts |
| 500 | ±2.2 pts | ±4.4 pts |
That is the optimistic case: it assumes independent trades and unchanging conditions. Neither holds cleanly, as how many trades before you can judge a strategy covers in detail.
So after a hundred trades at an observed 45%, a true win rate anywhere from the mid-30s to the mid-50s is consistent with what you saw. Now feed that range into Kelly.
Kelly amplifies the error you already have
Example: a strategy paying 2R per win.
| Win rate | Expectancy | Full Kelly fraction |
|---|---|---|
| 50% | +0.50R | 25.0% |
| 45% | +0.35R | 17.5% |
| 40% | +0.20R | 10.0% |
| 35% | +0.05R | 2.5% |
| 33% | −0.01R | negative — no edge |
A ten-point error in the win rate — well inside the range a hundred trades allows — is the difference between 17.5% and 2.5%. A seven-fold change in the output from an input error you could not have ruled out.
The derivative shows why: df/dp = (b + 1) ÷ b, so each point of error in p
moves the fraction by more than a point — and since the fraction is often small,
the relative error in the output dwarfs the relative error in the input.
The growth curve is not symmetric
If errors in p cut both ways equally, you might argue they average out. They
do not, because the cost of betting too much and betting too little is not the
same.
Long-run growth per trade is g(f) = p·ln(1 + b·f) + q·ln(1 − f). Continuing
the 45%-at-2R example, here is growth at multiples of the Kelly fraction,
relative to the maximum:
| Multiple of Kelly | Fraction | Growth vs maximum |
|---|---|---|
| 0.25× | 4.4% | 45% |
| 0.5× | 8.75% | 76% |
| 0.75× | 13.1% | 94% |
| 1× | 17.5% | 100% |
| 1.5× | 26.25% | 77% |
| 2× | 35.0% | 6% |
| 2.5× | 43.75% | negative |
Half Kelly keeps about three-quarters of the growth. Double Kelly keeps almost none, and a little beyond that the same positive-expectancy strategy shrinks the account. Undershooting costs you some growth; overshooting can cost you all of it.
Now combine the two sections. Suppose the estimate says 45% but the truth is 40%. The true Kelly fraction is 10%, so betting the estimated 17.5% is 1.75× the true optimum — well onto the steep side. Under that true 40% win rate, betting the estimated full Kelly gets under half the achievable growth, while betting half the estimate lands almost exactly on the true optimum. If the truth is 35%, full estimated Kelly has negative growth, half is slightly negative, and a quarter is barely positive.
Uncertainty plus asymmetry is the whole argument for fractional Kelly. It is not a softening of the maths; it is the maths applied to inputs you actually have.
Your estimate is probably biased upward, too
Sampling error is symmetric in principle. The process that produces your estimate is not.
- You chose the strategy because its numbers looked good. Of several approaches tested, the one kept is the one whose sample happened to run hot. That selection biases the surviving estimate upward.
- Backtests leak. Parameters tuned on the same data they are scored on, costs understated, slippage ignored.
- Win rate and payoff are both estimated. Kelly is sensitive to
bas well asp, and realised R tends to come in below planned R once fees, partial fills and early exits are counted.
So the error is not just wide — it leans toward overestimating the edge, which is the direction the growth curve punishes.
Full Kelly’s drawdowns are hard to survive
Even with a perfectly known edge, full Kelly is a rough ride. A well-known continuous-time result says that at full Kelly the probability of the balance ever falling to half its starting value is about 50%; at half Kelly it is about 12.5%.
A simulation of the 45%-at-2R example (10,000 runs of 200 trades each, with the edge known exactly) agreed closely:
| Sizing | Runs that fell to half the starting balance | Median worst peak-to-trough drawdown |
|---|---|---|
| Full Kelly (17.5%) | 46% | 90% |
| Half Kelly (8.75%) | 12% | 62% |
| Quarter Kelly (4.4%) | under 1% | 36% |
Every full-Kelly run had a drawdown of at least 50% somewhere in its 200 trades. Run the recoveries those imply through the drawdown recovery calculator — after a 90% drawdown the balance has to grow tenfold just to get back to its peak.
These are the figures when the edge is exactly right; estimation error makes them worse. And as risk of ruin points out, the drawdown that ends most accounts is the one where the trader stops following the rules, which usually arrives long before the arithmetic floor.
Trading breaks the other assumptions too
Kelly assumes sequential, independent bets with a fixed binary payoff. Trading departs from all three:
- Positions overlap. Two open trades on correlated instruments are closer to one larger bet; per-trade Kelly does not see that. See correlation and position limits.
- Losses cluster. The conditions producing one loss tend to produce the next, so losing streaks run longer than independent-trial arithmetic suggests.
- Payoffs are a distribution. Gaps through stops produce losses larger than 1R.
Each pushes the true optimum below the computed one.
What this means in practice
Kelly is still useful, as a ceiling rather than a setting:
- Compute it on a pessimistic win rate, not your point estimate. If your sizing exceeds Kelly on the low end of your plausible range, you are likely overbetting.
- Read the sensitivity, not the headline. If the fraction swings wildly across your uncertainty band, the edge is not known well enough to size from.
- If Kelly goes negative on a plausible input, that is information about the strategy, not about sizing.
- Pick your risk per trade from what you can survive — expected losing streaks and your real drawdown tolerance — then check it against Kelly, not the other way round. How to calculate position size covers turning that choice into a quantity.
FAQ
Why is full Kelly considered too aggressive?
Because it is optimal only when the win rate and payoff are known exactly, and in trading both are estimates from limited samples. The growth curve falls off steeply when you bet above the true optimum and gently when you bet below it, so an overestimated edge costs far more than an underestimated one. Fractional Kelly is the response to that uncertainty.
What is half Kelly?
Betting half the fraction the Kelly formula produces. With an exactly known edge it keeps roughly three-quarters of the maximum long-run growth while substantially reducing drawdowns, and when the edge has been overestimated it can land closer to the true optimum than full Kelly does. It is a response to input uncertainty, not a different formula.
How many trades do I need before Kelly is reliable?
More than most traders have. After 100 trades the standard error on a win rate near 45% is about five percentage points, and at 2R a ten-point error changes the Kelly fraction several-fold. Even large samples do not remove selection bias or changing market conditions.
Can the Kelly fraction be negative?
Yes. A negative result means the inputs imply no edge — expectancy is zero or below — and the formula’s answer is not to take the bet. If a plausible, slightly lower win rate produces a negative fraction, the edge is not well enough established to size from.