Position sizing is the most important decision in trading and the one most traders get wrong. They spend weeks optimizing entry signals and minutes deciding how much to bet. The entry determines direction. The position size determines whether you survive.
The Kelly criterion provides a mathematical answer to the position sizing question. Given your win rate and your average win-to-loss ratio, it tells you the fraction of capital to risk on each trade to maximize long-term growth. The formula was developed by John Kelly at Bell Labs in 1956 for information theory and has been applied to gambling and trading ever since. It is elegant, powerful, and will blow up your account if you use it naively in crypto.
The Formula
The Kelly fraction is calculated as: f equals p times b minus q, divided by b. In this formula, p is your win rate as a decimal between 0 and 1, q is your loss rate (1 minus p), and b is the payoff ratio, which is your average winning trade divided by your average losing trade.
If your strategy wins 55 percent of the time (p equals 0.55, q equals 0.45) and your average win is 1.5 times your average loss (b equals 1.5), the Kelly fraction is: (0.55 times 1.5 minus 0.45) divided by 1.5, which equals 0.25. The formula says to risk 25 percent of your capital on each trade to maximize long-term geometric growth.
This is mathematically optimal for maximizing the growth rate of your bankroll over infinite trades. It is also aggressive enough to produce drawdowns that no human can tolerate and no risk framework should allow.
Why Full Kelly Destroys Crypto Traders
The Kelly formula assumes your estimates of win rate and payoff ratio are exact. In practice, they are estimates derived from backtests. Your backtest might show a 55 percent win rate, but the true win rate in live conditions might be 50 percent. That five-point error changes the optimal Kelly fraction significantly.
The formula also assumes each trade is independent. In crypto, trades are often correlated because market conditions persist. A bearish regime produces correlated losses across multiple trades. The Kelly formula does not account for this serial dependence, and its recommended sizing is too aggressive when losses cluster.
Most importantly, the Kelly fraction maximizes growth rate but tolerates enormous drawdowns. A full Kelly bettor will experience a 50 percent drawdown approximately once every few hundred trades. In crypto, where volatility compresses this timeline, full Kelly drawdowns arrive faster and hit harder than in traditional markets.
Half-Kelly: The Production Solution
We use half-Kelly as our sizing baseline. The implementation multiplies the Kelly fraction by 0.5 before applying it. This is a standard practice across quantitative trading firms because it preserves roughly 75 percent of the growth rate while cutting the variance nearly in half.
The math works because the Kelly growth function is relatively flat near the optimum. Betting half the Kelly fraction gives up only about 25 percent of the theoretical maximum growth rate while dramatically reducing the probability and magnitude of large drawdowns. Betting twice the Kelly fraction gives the same growth rate as half-Kelly but with vastly more variance. The asymmetry strongly favors undersizing.
Our implementation handles several edge cases. If the win rate is zero, Kelly returns zero (do not trade). If the win rate is 100 percent, it caps at 0.5 (half-Kelly maximum). If the Kelly fraction is negative (the strategy has negative expected value), it returns zero. A negative Kelly means the edge does not exist and no amount of sizing optimization can create one.
Risk Parity Allocation Across Bots
Kelly sizing determines how much to risk per trade within a single bot. But we run 45 bots simultaneously, and each bot needs a capital allocation. This is a different problem: how to distribute capital across multiple strategies with different risk characteristics.
We use inverse-volatility weighting, also called risk parity. Bots with lower volatility receive more capital. Bots with higher volatility receive less. The logic is that equal dollar allocation to strategies with different volatilities creates unequal risk contribution. A high-volatility strategy dominates the portfolio's risk even if it has the same dollar allocation as a low-volatility strategy.
The formula calculates inverse volatility for each bot (1 divided by the bot's recent return volatility), sums all inverse volatilities, and allocates proportionally. A single-bot cap of 40 percent prevents any one bot from dominating the portfolio regardless of how low its volatility is.
When the cap clips a bot's allocation, the surplus capital is redistributed to uncapped bots using the same inverse-volatility weights. This ensures all capital is deployed and the portfolio maintains its risk parity characteristic even after concentration limits are applied.
Position Sizing in Practice
Our production position sizing works in three layers. First, the bot receives its capital allocation from the risk parity allocator. Currently this is 1,000 dollars per bot (equal allocation since we are in paper trading phase, with risk parity to be activated for live trading).
Second, the per-trade risk is set at 10 percent of the bot's current equity. This means a bot with 1,000 dollars risks 100 dollars per trade. As the bot makes money, the 10 percent scales up. As it loses money, the 10 percent scales down. This is essentially a fixed-fractional approach that naturally compounds on the upside and de-risks on the downside.
Third, the correlation sizer adjusts the position based on how correlated the new trade is with existing portfolio positions. If the portfolio already has significant long exposure to altcoins and the new signal is another altcoin long, the correlation sizer reduces the position to between 25 and 100 percent of its intended size. More on this in a later post.
The per-bot risk checks then validate the sized position. Maximum position size is 25 percent of bot capital (250 dollars on a 1,000 dollar bot). If the sized position exceeds this, it is clipped. The consecutive loss cooldown can further reduce sizing to as low as 10 percent of normal after extended losing streaks.
Kelly for Strategy Selection
Beyond position sizing, the Kelly criterion is useful for strategy evaluation. We calculate the Kelly fraction for every strategy that passes our validation pipeline. A strategy with a higher Kelly fraction has a larger mathematical edge.
Our mean reversion strategy on Bollinger Bands has the highest Kelly fractions across our strategy lineup, which aligns with its high validated Sharpe ratios (9 to 19 across symbols). The 4-hour momentum strategy on BTC and ETH has lower Kelly fractions (Sharpe 1.7 to 3.9), which means smaller optimal bet sizes. This automatically translates into appropriate position sizing: stronger edges get larger bets, weaker edges get smaller bets.
We display the Kelly fraction on strategy risk scorecards in the dashboard. This gives an immediate sense of a strategy's sizing potential. A Kelly fraction of 0.15 means the strategy supports up to 7.5 percent allocation per trade at half-Kelly. A Kelly fraction of 0.05 means only 2.5 percent per trade. These numbers directly inform capital allocation decisions.
Common Kelly Mistakes
The most common mistake is using full Kelly. No institutional quant fund uses full Kelly for the reasons described above. Half-Kelly or even quarter-Kelly is standard practice. The growth sacrifice is small and the risk reduction is enormous.
The second mistake is calculating Kelly from a single backtest period. Your win rate and payoff ratio are period-specific. A strategy that wins 60 percent in a bull market and 45 percent in a bear market does not have a 60 percent win rate. You need the win rate across all market conditions, which is why we calculate Kelly after regime validation across five distinct market periods.
The third mistake is ignoring transaction costs. The Kelly formula operates on net returns. If your gross win rate is 55 percent but fees reduce it to 52 percent, the Kelly fraction changes significantly. Our backtest engine applies realistic fees (0.10 percent taker on each side) and slippage (2 to 10 basis points) before calculating any metrics, so our Kelly fractions reflect real-world conditions.
The fourth mistake is applying Kelly to strategies with negative expected value. If the Kelly fraction is negative, the correct position size is zero. No amount of sizing optimization can make a negative-edge strategy profitable. This is why we calculate Kelly only after validation, not after initial backtesting. A sweep Sharpe of 1.98 can correspond to a negative out-of-sample edge, as we learned with the Ornstein-Uhlenbeck strategy.
The Sizing Stack
In summary, our position sizing works as a stack: risk parity allocates capital across bots, half-Kelly determines the base risk per trade, the correlation sizer adjusts for portfolio overlap, and per-bot risk gates enforce hard limits. Each layer addresses a different dimension of the sizing problem. No single formula handles all of them. The Kelly criterion is the mathematical foundation, but it is one component in a system that recognizes the difference between theory and production.