Every trader eventually faces the question that matters more than which direction to trade: how much to risk. Get the direction wrong and you lose one trade. Get the sizing wrong and you lose the account. The Kelly criterion provides the mathematical answer to position sizing, and it is simultaneously the most powerful and the most dangerous tool in quantitative trading.
The Formula and What It Means
The Kelly fraction is defined as f* = (p * b - q) / b, where p is the probability of winning (win rate), b is the ratio of average win to average loss (payoff ratio), and q is the probability of losing (1 minus p).
Suppose a strategy wins 58 percent of the time with an average win of 1.3 times the average loss. The Kelly fraction is (0.58 * 1.3 - 0.42) / 1.3, which equals (0.754 - 0.42) / 1.3, which equals 0.257. The formula recommends risking 25.7 percent of capital on each trade.
This fraction maximizes the long-run geometric growth rate of the bankroll. Over a theoretically infinite number of trades, no other fixed-fraction strategy will compound wealth faster. That theoretical optimality is exactly why it is dangerous in practice.
Why Full Kelly Is Unsurvivable
The Kelly fraction assumes perfect knowledge of the true win rate and payoff ratio. In practice, we estimate these from backtests. A backtest might show a 58 percent win rate, but the true live win rate might be 53 percent due to slippage, data discrepancies, or regime changes. A five-point overestimate of the win rate can shift the Kelly fraction from 0.257 to something much smaller, meaning you are actually overbetting relative to your true edge.
Full Kelly also tolerates drawdowns that no risk framework should allow. A full Kelly bettor will experience a 50 percent drawdown with near certainty over a long enough horizon. In crypto, where volatility compresses timelines, that drawdown arrives faster than in equities. A strategy that is correctly sized at full Kelly can lose half its capital in a week during a regime change, even if the strategy is profitable in the long run.
The mathematical insight is that the Kelly growth curve is flat near the peak. Betting half the Kelly fraction sacrifices only about 25 percent of the theoretical growth rate while cutting variance nearly in half. Betting double the Kelly fraction gives the same growth rate as half-Kelly but with enormously more volatility. The asymmetry is stark and universally favors undersizing.
Half-Kelly: Our Production Standard
We multiply the Kelly fraction by 0.5 before applying it to any bot. This is not a conservative choice made out of caution. It is the standard practice across quantitative firms because the math strongly supports it. The reduction in growth rate is minor. The reduction in drawdown probability is substantial.
Our implementation includes several edge-case guards that matter in production. When the win rate is zero, the formula returns zero. There is no trade to make if you never win. When the win rate is 1.0 (which can happen in small backtests with few trades), the fraction caps at 0.5 rather than recommending full capital deployment. No backtest sample is large enough to justify 100 percent conviction.
When the Kelly fraction is negative, it means the strategy has negative expected value. The formula returns zero. A negative Kelly is not a signal to go short or to reduce position size. It is a signal that the strategy should not be traded at all. No position sizing can make a negative-edge strategy profitable.
Real Kelly Fractions From Our Strategies
Our backtest validation runs produce Kelly fractions for every strategy-symbol combination. These numbers reveal important differences between strategies.
Mean reversion on Bollinger Bands (bb_period=30, bb_std=2.5) on high-beta altcoins produces Kelly fractions between 0.15 and 0.30 depending on the symbol. PEPE and WIF tend toward the higher end because their volatility creates larger win/loss ratios. After the half-Kelly multiplier, these become 0.075 to 0.15, meaning each bot risks 7.5 to 15 percent of its allocated capital per trade.
Momentum RSI+MACD on 15-minute timeframes produces lower Kelly fractions, typically 0.08 to 0.18 before the half-Kelly adjustment. Momentum strategies have slightly lower win rates than mean reversion but compensate with better payoff ratios on trending moves. After half-Kelly, the risk per trade is 4 to 9 percent of capital.
The 4-hour momentum strategy on BTC and ETH produces the lowest Kelly fractions in our portfolio, around 0.06 to 0.12. These are less volatile assets on longer timeframes, so the edge per trade is smaller. Half-Kelly brings these to 3 to 6 percent risk per trade.
Derivatives-based strategies like leverage_composite produce Kelly fractions of 0.10 to 0.20. The funding rate and open interest signals provide a genuine but modest edge. These strategies trade less frequently, and each trade carries more conviction, which is reflected in higher payoff ratios despite moderate win rates.
From Kelly Fraction to Position Size
The Kelly fraction tells you what fraction of capital to risk. Converting this to a position size requires knowing your stop-loss distance. If your Kelly fraction (after half-Kelly) is 0.10 and your stop-loss is 3 percent away from entry, the position size is 0.10 divided by 0.03, which equals 3.33 times your capital allocation.
This means leverage. In crypto perpetual futures, leverage is readily available, which makes Kelly sizing both more practical and more dangerous. The Kelly fraction inherently accounts for the probability and magnitude of losses, so the leverage it implies is theoretically appropriate. But in practice, exchange liquidation mechanics, funding costs, and fat-tailed price distributions make theoretical Kelly leverage too aggressive.
Our risk framework adds another constraint on top of Kelly: no single position can exceed 1x the bot's allocated capital (no leverage in paper trading phase). This means the effective risk per trade is often smaller than what Kelly recommends because the position size cap binds before the Kelly fraction does.
The 40 Percent Single-Bot Cap
Kelly and half-Kelly address per-trade sizing within a single bot. But capital allocation across 45 bots is a separate problem. Our risk parity allocator distributes capital using inverse-volatility weighting, with a hard cap of 40 percent on any single bot.
The cap exists because inverse-volatility weighting can produce extreme allocations. If one bot trades a low-volatility strategy while 44 others trade high-volatility strategies, the math might assign 60 or 70 percent of capital to the one low-vol bot. This defeats the purpose of running 45 bots. The 40 percent cap ensures no single bot dominates the portfolio regardless of its volatility characteristics.
When the cap clips a bot's allocation, the surplus capital is redistributed to uncapped bots using the same inverse-volatility weights. Every dollar gets deployed. The portfolio maintains its risk parity character within the constraint that no bot gets more than 40 percent.
Kelly in Context
The Kelly criterion is a tool, not a religion. It provides the mathematical framework for thinking about position sizing, but it requires adjustments for the real world: half-Kelly for estimation uncertainty, edge-case guards for degenerate inputs, position caps for leverage control, and portfolio-level allocation constraints for multi-bot environments.
The most important lesson from Kelly is directional, not numerical. If your edge is small, your position should be small. If your edge is negative, your position should be zero. Many traders size their positions based on conviction or excitement rather than expected value. Kelly forces the discipline of sizing proportionally to the actual measured edge, not the hoped-for edge.