Half Kelly vs Full Kelly When Your Edge Is a Guess

You backtested a setup, the numbers look good, and a Kelly calculator tells you to risk 8% of your account per trade. That feels wrong, and it is. Kelly is the size that maximizes growth if your edge is known exactly, and your edge is never

You backtested a setup, the numbers look good, and a Kelly calculator tells you to risk 8% of your account per trade. That feels wrong, and it is. Kelly is the size that maximizes growth if your edge is known exactly, and your edge is never known exactly, so the real question is how far below Kelly to stay.

Kelly is a growth-maximizing size, not a safe size

The Kelly criterion comes from John Kelly's 1956 paper, A New Interpretation of Information Rate, published in the Bell System Technical Journal. For a repeated bet with a known edge, it gives the fraction of your bankroll that maximizes the long-run growth rate of capital.

For a trading strategy the formula is f = p − (1 − p) / b. Here p is your win rate and b is your payoff ratio, meaning average win divided by average loss. The result f is the share of equity you lose when a trade hits its stop, so it is the same thing as risk per trade.

Take a strategy that wins 45% of the time and pays 1.5R on winners (R is your initial risk on a trade). Its expectancy (average result per trade) is +0.125R. Plug it in: 0.45 − 0.55 / 1.5 = 0.083. Full Kelly says risk 8.3% of equity per trade.

Our position sizing pillar guide covers where Kelly sits among the other sizing tools. This post goes one step further and puts numbers on what happens when you move away from the exact Kelly size in either direction.

Overbetting costs more than underbetting

Growth as a function of bet size is a hill, and it is not symmetric in what it does to you. Edward Thorp, who ran Kelly sizing in blackjack and markets, states the practical point plainly in his chapter on the Kelly criterion: using too large a fraction and overbetting is much more severely penalized than using too small a fraction and underbetting.

The same chapter works out the risk side in a continuous-time model. Half Kelly has 3/4 of the growth rate of full Kelly. The chance of ever being cut to half your starting capital is 1/2 at full Kelly and 1/8 at half Kelly.

We checked that against a simulation. Same strategy as above, 45% wins, 1.5R winners, no costs, independent trades, and a known edge. We ran 100,000 sequences of 500 trades at each size.

Our numbers sit slightly under Thorp's because we stopped at 500 trades, and the model is a continuous approximation. The shape is the same.

Risk per trade	Median equity after 500 trades	Median max drawdown	Finished below start
0.25 × Kelly (2.1%)	3.1×	31%	2.7%
0.5 × Kelly (4.2%)	6.9×	54%	4.8%
1.0 × Kelly (8.3%)	13.1×	83%	13.1%
1.5 × Kelly (12.5%)	7.0×	96%	28.1%
2.0 × Kelly (16.7%)	1.1×	99%	48.3%

Read the 0.5 and 1.5 rows side by side. They grow at almost the same median rate, yet the 1.5 row carries a 96% median drawdown against 54%. That is the overbetting penalty in one line. Past full Kelly you pay more risk for less growth, and at 2× the growth is gone entirely.

A 13× median result and an 83% drawdown also tell you something about what full Kelly feels like. Even with a perfectly known edge, the typical full Kelly path loses most of the account at some point on the way up. Very few people hold through that.

Your Kelly number is an estimate with an error bar

Everything above assumed you know p and b. You know them only from a sample of trades, and the sample has noise. The standard error of a win rate is the square root of p(1 − p) / n, so at p = 45% it is about 5 percentage points with 100 trades. That is the same arithmetic behind how many trades a backtest needs before the win rate means anything.

Five points sounds small. Through the Kelly formula it is not. With 1.5R winners, the breakeven win rate is 40%, so a strategy with a true 45% win rate sits only five points above zero edge. A one standard deviation miss on the win rate moves Kelly by about its own full value.

We simulated that. The true strategy is the same 45% / 1.5R one, so true Kelly is 8.3%. For each trial we draw an estimated win rate from a track record of n trades, compute Kelly from that estimate, and record how it compares to the truth.

Trades in the track record	Estimate says no edge (win rate ≤ 40%)	Estimated Kelly is at least 2× true Kelly
50	19.8%	28.3%
100	13.5%	18.2%
200	6.7%	8.9%
500	1.1%	1.4%

At 100 trades, nearly one in five Kelly calculations lands at double the true size or more. In Thorp's continuous model, growth is a parabola in bet size that falls back to zero at 2× Kelly (with no return on idle cash), and our growth figure for the 2× row agrees. The calculator is not wrong, the input is. It is the same effect that makes a backtest's max drawdown one draw from a wide range, applied to your win rate.

Sizing off an estimate makes half Kelly the better bet

Now the part that decides the question. We gave each simulated trader a 100-trade track record, had them compute Kelly from their estimated win rate (betting nothing if the estimate showed no edge), and then traded 500 live trades with the true 45% / 1.5R strategy at a fraction of that estimated Kelly.

Fraction of estimated Kelly	Median equity after 500 trades	Finished below start	Finished below half of start	Drawdown of 50% or more at some point
0.25	2.3×	2.7%	0.5%	23%
0.5	3.4×	6.8%	3.6%	53%
1.0	2.3×	21.9%	18.1%	70%

With a known edge, full Kelly had the highest median result. With an estimated edge, half Kelly beats it on median growth, 3.4× against 2.3×, and has about a third of the chance of finishing below the starting balance. The people who overshot, and about 18% of them did, dragged the full Kelly column down.

This is why Thorp argues for fractional Kelly when estimates are uncertain. In the same chapter he notes that if your estimated edge is uncertain, it is wise to assume the true edge is lower than the estimate and to choose a fraction low enough to prevent growth from turning negative. For further reading on both sides, MacLean, Thorp and Ziemba's paper on the good and bad properties of Kelly and fractional Kelly is the standard academic reference.

Kelly tells you how fast you could grow if you knew your edge. It says nothing about how well you know it.

— The Tradoki desk note

What to do with the number

A practical reading, as a rule of thumb and not a measured statistic: treat full Kelly as a ceiling you should not approach, and compute it from the conservative end of your estimate, not the midpoint. If a backtest suggests a 45% win rate over 100 trades, run the formula at 40% as well and see what the answer says about whether the edge exists at all.

Two other checks matter more than the fraction itself. First, drawdown: even at a quarter of Kelly in our known-edge run, the median max drawdown was 31%, so the size has to fit the loss you can survive and the rules you trade under. A prop firm limit like a trailing drawdown that resets intraday or at end of day cuts the usable size far below any Kelly figure.

Second, the Kelly model assumes independent trades with fixed odds. Real markets cluster volatility, costs and slippage eat the edge, and stops get filled worse than planned. Those all push the true edge below the backtest, which is a reason to add margin, never to remove it.

● FAQ

What is the Kelly criterion formula for trading?

For a strategy with win probability p and an average win of b times the average loss, the Kelly fraction of equity to put at risk is p minus (1 minus p) divided by b. A 45% win rate with 1.5R winners gives 0.45 minus 0.55 divided by 1.5, about 8.3% of equity per trade. That number is the growth-maximizing size only if p and b are exactly right.

Why do people use half Kelly instead of full Kelly?

Half Kelly keeps about 75% of full Kelly's long-run growth rate while cutting the chance of a deep loss sharply. Thorp's analysis puts the chance of ever losing half your starting capital at 1/2 for full Kelly and 1/8 for half Kelly. The bigger reason in practice is that your edge is estimated, and overestimating it is costly.

What happens if I bet more than full Kelly?

Growth falls, and risk keeps rising. In our simulation, betting 1.5 times Kelly produced about the same median growth as betting half Kelly, with a median max drawdown of 96% instead of 54%. At 2 times Kelly the long-run growth rate is roughly zero.

Can I use Kelly if I only have 100 trades of data?

You can compute it, but the answer is a wide range, not a number. With a true 45% win rate and 1.5R winners, an estimate from 100 trades overshot twice the true Kelly size about 18% of the time. Use the number as a ceiling on risk per trade, and size well below it.

Is Kelly the same as risking 1% per trade?

No. Kelly is derived from your win rate and payoff, so for a decent edge it often lands far above 1%. A trader risking a fixed 1% on a modest edge is usually at a small fraction of Kelly, which is a different and safer place to be.