Kelly Criterion trading uses a simple formula, f* = W - (1-W)/R, to calculate the fraction of your account you should risk on each trade. W is your win rate as a decimal. R is your average win divided by your average loss. The output is the position size that maximizes long-term geometric growth, assuming your inputs are accurate. In practice, almost no serious trader runs Full Kelly. They run Half or Quarter Kelly because W and R are estimates, and estimation error kills accounts faster than bad trades.
This guide walks through the math, three worked examples pulled from real trade data, and the correlation and drawdown problems that make Full Kelly dangerous in leveraged markets like crypto futures.
What is the Kelly Criterion in trading?
The Kelly Criterion is a position sizing formula developed by John L. Kelly Jr. at Bell Labs in 1956. It was originally built to handle noise in long-distance phone signals, but Ed Thorp adapted it for blackjack and later for the stock market. Today it sits at the core of how quantitative funds size bets when they have a genuine edge.
The formula answers one question: given a repeated bet with a known probability of winning and a known payoff ratio, what fraction of my capital should I stake on each bet to grow my bankroll as fast as possible without going broke?
The answer is not intuitive. Bet too little and you leave growth on the table. Bet too much and volatility compounds against you until your equity curve collapses. Kelly finds the exact fraction that maximizes the geometric growth rate, not the arithmetic one. That distinction is why Kelly is the correct framework for compounding traders and why simple expected value calculations underweight the danger of ruin.
The Kelly Criterion formula
The standard formula for a binary bet is:
f = W - (1 - W) / R*
Where:
- f* is the optimal fraction of your capital to risk on the trade
- W is your win rate, expressed as a decimal (e.g. 0.55 for 55%)
- R is your reward-to-risk ratio, calculated as average win size divided by average loss size
A quick sanity check. If W = 0.5 and R = 1 (a fair coin flip with 1:1 payoff), then f* = 0.5 - 0.5/1 = 0. Kelly correctly tells you not to bet when there is no edge. If your edge disappears, your optimal size is zero.
Worked example 1: a swing trader with a 55% hit rate
Suppose your last 200 swing trades show a 55% win rate. Your average winning trade returned 1.8R (where R is the risk you took on entry, defined by your stop). Your average losing trade lost 1.0R because you honored your stops.
- W = 0.55
- R = 1.8 / 1.0 = 1.8
f* = 0.55 - (1 - 0.55) / 1.8 = 0.55 - 0.25 = 0.30, or 30%.
Full Kelly says risk 30% of your account per trade. That number is terrifying, and it should be. It is also the reason nobody runs Full Kelly in practice. A single normal losing streak of four trades at 30% each takes you from $10,000 to $2,401. Half Kelly (15%) is more sensible but still aggressive. Quarter Kelly (7.5%) is closer to what a professional would actually use, and even that is high for most retail traders.
Worked example 2: a scalper with a high hit rate and small edge
Scalpers often win 70%+ of trades but with a poor reward-to-risk. Say W = 0.72 and average win/loss = 0.5R/1.0R, so R = 0.5.
f* = 0.72 - (1 - 0.72) / 0.5 = 0.72 - 0.56 = 0.16, or 16%.
Even with a 72% hit rate, Kelly says 16% per trade because the payoff is asymmetric against you. If R drops to 0.4 (winners get even smaller), f* = 0.72 - 0.28/0.4 = 0.02, essentially zero. This is why scalping strategies collapse the moment slippage, funding, or fees eat into R. The edge is fragile and the correct position size shrinks fast.
Worked example 3: a crypto futures trader with a losing hit rate
Many trend followers win only 35% of the time but let winners run. Say W = 0.35 and R = 3.5 (average win is 3.5x average loss).
f* = 0.35 - (1 - 0.35) / 3.5 = 0.35 - 0.186 = 0.164, or 16.4%.
A losing hit rate with a strong R still produces a healthy Kelly fraction. This is the mathematical reason trend-following works over decades even though it feels miserable trade by trade. Kelly rewards asymmetric payoffs, not high win rates.
Why traders use Half Kelly or Quarter Kelly
Full Kelly assumes your W and R are the true, long-run parameters. In trading they never are. They are estimates from a finite sample, and small errors in the inputs produce large errors in the output.
Here is the uncomfortable math. If your true edge gives a Full Kelly of 20% but your journal overestimates W by just 3 percentage points, your "optimal" 20% is actually closer to 8%. You will now be overbetting by 2.5x, and your drawdowns will be brutal. Kelly is asymmetric on the downside: overbetting punishes you geometrically, underbetting only costs linear growth.
That is why the standard practice is:
- Half Kelly: roughly 75% of Full Kelly's expected growth with about 50% of the volatility
- Quarter Kelly: roughly 50% of the growth with dramatically lower drawdowns and higher psychological survivability
Ed Thorp, who ran Kelly in real markets for decades, publicly recommended fractional Kelly for exactly this reason. If you cannot stomach a 50% drawdown, Full Kelly is not for you, and 50% drawdowns are the expected behavior of Full Kelly systems, not the tail case.
How to extract W and R from your trade journal
The formula is trivial. Getting reliable inputs is the hard part. Here is the workflow:
- Filter by strategy. Never compute Kelly across all your trades mixed together. A breakout strategy and a mean-reversion strategy have different W and R, and averaging them is meaningless. Segment first.
- Use R-multiples, not dollars. Convert every closed trade to a multiple of the risk you took on entry. A trade that made 2.3x your initial stop distance is +2.3R. This normalizes across position sizes and account growth.
- Compute W as wins / (wins + losses). Break-even trades usually get excluded. Be consistent.
- Compute R as average(win R-multiple) / average(loss R-multiple in absolute value). If your average win is +2.1R and your average loss is -0.95R, then R = 2.1 / 0.95 = 2.21.
- Require at least 50 trades per strategy, ideally 100+. Kelly outputs from 20-trade samples are noise. Wide confidence intervals mean you should scale your Kelly fraction down further.
- Recompute quarterly. Edges decay. Your W and R from 2022 crypto are not your W and R today.
Most traders skip step 1 and step 2 entirely, which is why their Kelly outputs are unreliable. Segmentation and R-normalization are non-negotiable if you want the formula to mean anything.
How TraderNest helps you calculate Kelly correctly
Running Kelly by hand is easy. Keeping your W and R honest across dozens of trades, multiple strategies, and several exchanges is not. TraderNest auto-syncs trades from Bybit, Binance, OKX, Bitget, MEXC, KuCoin, Gate.io, Kraken, Deribit, and Hyperliquid, so every closed position lands in your journal with its actual entry, exit, and P&L. No manual typing, no forgotten trades skewing the sample.
From there, the R/R analysis and strategy analysis pages compute your average win, average loss, win rate, and reward-to-risk per strategy. That gives you clean inputs for the Kelly formula per system, not a meaningless blended average.
The deeper problem is behavioral. Kelly assumes you actually take the trades your strategy signals and honor your stops. If you cut winners early or move stops, your realized R collapses and your "optimal" size becomes overbetting. AI Hawk detects exactly these patterns automatically: Premature Exits, Inconsistent Risk Management, Plan Discipline breaches, and Post-Loss Confidence issues. If Hawk flags that you exit winners 40% earlier than your plan, your true R is lower than your journal suggests, and your Kelly fraction should shrink accordingly.
Limitations and honest caveats
Kelly is a powerful framework but it is not a silver bullet. The main limitations:
- Estimation error dominates. Your W and R are guesses. Fractional Kelly exists because of this, not despite it.
- Trades are not independent. Kelly assumes each bet is independent of the last. In trending markets, correlated wins and losses cluster, and drawdowns exceed the theoretical model.
- Correlated positions compound risk. If you hold BTC, ETH, and SOL long simultaneously, you do not have three positions. You have one leveraged bet on crypto beta. Standard Kelly ignores correlation. For multi-position portfolios, you need correlation-adjusted Kelly, which scales each position down based on its correlation with the others.
- Payoff distributions are not binary. Real trades produce a distribution of outcomes, not a fixed win amount and fixed loss amount. The binary formula is an approximation. For continuous returns, use the log-optimal version: f* = mean_return / variance_of_returns.
- Kelly maximizes growth, not comfort. The path to maximum geometric growth passes through drawdowns that most traders quit before recovering from. If you cannot survive the drawdown emotionally, Kelly is mathematically correct and behaviorally wrong for you.
Practical checklist before you use Kelly
Before applying the formula to a live account:
- You have 100+ closed trades on the specific strategy
- Trades are normalized to R-multiples, not raw dollars
- W and R are computed per strategy, not blended
- You are using Half Kelly or Quarter Kelly, not Full
- You have modeled the expected drawdown and can survive it
- You recompute W and R at least quarterly
- You have separate Kelly fractions for uncorrelated strategies and cap total portfolio risk
Skip any of these and Kelly will feel like it is working right up until the drawdown that ends the account.
Ready to size positions like a professional?
Kelly is one tool inside a broader risk management framework covering position sizing, stop placement, correlation, and drawdown control. TraderNest gives you the clean data and behavioral pattern detection you need to trust your W and R before you bet on them. Read the full guide on risk management in trading to see how Kelly fits alongside R-multiples, fixed fractional sizing, and portfolio-level heat.
