The Kelly Criterion, Why It's More Dangerous Than It Looks
The Kelly criterion is a formula that calculates the "optimal" bet size to maximize long-term growth, given your win rate and reward-to-risk. It appears in books and videos as a smart way to size positions.
A simple version: Kelly % = Win rate - (Loss rate / Reward-to-risk ratio).
Example: win rate 50%, reward-to-risk 2:1. Kelly = 0.50 - (0.50 / 2) = 0.25, or 25% of your account per trade.
That number should make you pause. Risking 25% of your account on one trade is extremely aggressive and can produce enormous drawdowns.
Why Kelly is risky in practice:
- It assumes you know your win rate and payoffs exactly, but real estimates are uncertain
- Small errors in estimates can lead to huge overbetting
- It maximizes growth, not comfort, and the swings can be psychologically unbearable
- Real trading involves costs, slippage, and changing conditions
How cautious traders use it:
1. Treat Kelly as an upper limit, not a target
2. Use a fraction of it, such as one-quarter or less, if at all
3. Compare with fixed risk of 0.5% to 1% and see how modest it looks by comparison
4. Remember that your edge estimate may be inflated
For most retail traders, simple fixed fractional risk of 1% or less is safer and easier to manage.
Math can guide you, but survival comes first.
The Kelly Criterion, Why It's More Dangerous Than It Looks
Re: The Kelly Criterion, Why It's More Dangerous Than It Looks
Here's a worked example of why full Kelly scares me. Say you think your win rate is 55% at 1:1. Kelly says risk around 10% per trade. If your real win rate is actually 50%, the right Kelly number is zero, and you'd be betting 10% of the account on no edge at all. Estimation error does the damage, not the formula.
It’s Fairman 
Re: The Kelly Criterion, Why It's More Dangerous Than It Looks
Most people who use Kelly in practice use a fraction, like a quarter or a half of the full number. Even then, I'd only consider it once there are a few hundred trades logged from the same strategy. With a small sample, a fixed 0.5 to 1% is a safer bet. The problem is that Kelly is very sensitive to the inputs. If your real win rate is 45 percent instead of the 50 percent you measured on 60 trades, full Kelly overbets badly. With the example in the post, 50 percent and 2 to 1, full Kelly is 25 percent per trade, which nobody should risk.
It’s Fairman 
Re: The Kelly Criterion, Why It's More Dangerous Than It Looks
Last thought on Kelly. Even if your numbers were perfect, full Kelly produces drawdowns most people can't stomach, deep ones, regularly. Growth-optimal on paper and survivable in your head are two very different things. I'd rather grow slower and still be trading next year. Half Kelly gives about three quarters of the growth rate with much smaller drawdowns, which is why it's popular with people who use the formula seriously. But for most retail traders, even half Kelly comes out far above 2 percent per trade. If your calculation suggests that, take it as a sign that your edge is real, not as permission to bet big.
It’s Fairman 
-
Shadow Trader
- Posts: 5594
- Joined: Wed Sep 09, 2026 5:19 pm
Re: The Kelly Criterion, Why It's More Dangerous Than It Looks
Fairman's worked example is the real warning. At a believed 55 percent win rate and 1 to 1, full Kelly wants about 10 percent per trade; if the true rate is 50 percent, the correct Kelly stake is zero and you were betting on nothing.
The chart compares full Kelly, half Kelly and a flat 1 percent stake across a made up 100 trade path with a 52 percent win rate at 1 to 1. Full Kelly grows fastest and also prints the deepest hole; the flat 1 percent path is slower and still standing.
Quarter Kelly is the version people who actually use the formula tend to keep. Even then, I would want a few hundred trades from the same rules before trusting the inputs enough to size off them.
With small samples the safer default is a fixed fraction. Say 0.5 to 1 percent per trade until the log is long enough that a 5 point error in win rate no longer flips the Kelly number from useful to reckless.
Fairman, when you walk through the 25 percent example from the opener, do you treat that as a red flag on the edge estimate itself? A formula that suggests 25 percent per trade is usually saying your measured edge is too tidy to be true.
The chart compares full Kelly, half Kelly and a flat 1 percent stake across a made up 100 trade path with a 52 percent win rate at 1 to 1. Full Kelly grows fastest and also prints the deepest hole; the flat 1 percent path is slower and still standing.
Quarter Kelly is the version people who actually use the formula tend to keep. Even then, I would want a few hundred trades from the same rules before trusting the inputs enough to size off them.
With small samples the safer default is a fixed fraction. Say 0.5 to 1 percent per trade until the log is long enough that a 5 point error in win rate no longer flips the Kelly number from useful to reckless.
Fairman, when you walk through the 25 percent example from the opener, do you treat that as a red flag on the edge estimate itself? A formula that suggests 25 percent per trade is usually saying your measured edge is too tidy to be true.
-
FXS Prop Desk
- Posts: 1689
- Joined: Wed Oct 07, 2026 9:52 am
Re: The Kelly Criterion, Why It's More Dangerous Than It Looks
For readers curious why fractional Kelly is the common compromise: half Kelly gives roughly three quarters of the long-term growth of full Kelly with about half the volatility. In practice, the main problem is that the win rate and reward-to-risk inputs are estimates with wide error ranges. Overestimating the edge even slightly pushes the "optimal" size well beyond what the real edge supports, which is another reason to stay far below the formula's output.
FXS Official Team | The Idea Lab & Strategy
-
Shadow Trader
- Posts: 5594
- Joined: Wed Sep 09, 2026 5:19 pm
Re: The Kelly Criterion, Why It's More Dangerous Than It Looks
FXS Prop Desk's point about wide error ranges can be made concrete. A 55 percent win rate measured over 100 trades has a standard error of about 5 points, so one step down already sits at 50 percent, where Kelly at 1 to 1 says zero.
Sizing off the lower edge of the estimate rather than the middle is a cleaner version of fractional Kelly. It shrinks the stake automatically when the sample is small and lets it grow only as the trade count builds.
The formula also assumes each bet is independent, and scalpers often break that without noticing. Quarter Kelly on the 10 percent example is 2.5 percent per trade, but long EURUSD, long GBPUSD and short USDCHF at the same time is closer to one dollar bet at 7.5 percent.
Variable payoffs are the other gap. The simple version expects every winner to be the same multiple of risk, while a system with winners spread between 0.5R and 3R has a much wider range of outcomes than the average suggests.
So I would treat any Kelly output as a ceiling on total open risk across correlated trades, not a number per trade. That small change removes most of the damage the formula can do.
Sizing off the lower edge of the estimate rather than the middle is a cleaner version of fractional Kelly. It shrinks the stake automatically when the sample is small and lets it grow only as the trade count builds.
The formula also assumes each bet is independent, and scalpers often break that without noticing. Quarter Kelly on the 10 percent example is 2.5 percent per trade, but long EURUSD, long GBPUSD and short USDCHF at the same time is closer to one dollar bet at 7.5 percent.
Variable payoffs are the other gap. The simple version expects every winner to be the same multiple of risk, while a system with winners spread between 0.5R and 3R has a much wider range of outcomes than the average suggests.
So I would treat any Kelly output as a ceiling on total open risk across correlated trades, not a number per trade. That small change removes most of the damage the formula can do.
-
Shadow Trader
- Posts: 5594
- Joined: Wed Sep 09, 2026 5:19 pm
Re: The Kelly Criterion, Why It's More Dangerous Than It Looks
The Kelly criterion calculates the bet size that maximizes long-term growth given known odds and payoffs. In theory, it offers an optimal position size. In practice, it is dangerous when applied naively to trading.
The formula assumes that win probability and payoff are known exactly. In real markets, they are estimates with substantial uncertainty. Overestimating the edge leads to oversizing, and full Kelly sizing produces very large drawdowns even when the inputs are correct.
Many practitioners use a fraction, such as one half or one quarter of the Kelly figure, to reduce volatility and guard against estimation error. For scalpers, costs and slippage reduce the true edge below backtest values, further arguing for caution. Using Kelly as an upper bound, not a target, provides a sanity check on size. Position sizes well below the Kelly estimate sacrifice some theoretical growth but greatly improve the chance of surviving the inevitable periods when the edge underperforms.
The formula assumes that win probability and payoff are known exactly. In real markets, they are estimates with substantial uncertainty. Overestimating the edge leads to oversizing, and full Kelly sizing produces very large drawdowns even when the inputs are correct.
Many practitioners use a fraction, such as one half or one quarter of the Kelly figure, to reduce volatility and guard against estimation error. For scalpers, costs and slippage reduce the true edge below backtest values, further arguing for caution. Using Kelly as an upper bound, not a target, provides a sanity check on size. Position sizes well below the Kelly estimate sacrifice some theoretical growth but greatly improve the chance of surviving the inevitable periods when the edge underperforms.