Position sizing decides how much of your account rides on each bot signal, and it matters more than signal accuracy for long-term survival. Two frameworks dominate the discussion: the Kelly criterion, which calculates the mathematically optimal fraction of capital to risk based on win rate and payoff, and fixed fractional risk, which risks the same small percentage of equity on every trade regardless of perceived edge. Both are respectable ideas, but they behave very differently when the inputs are uncertain, and in live trading the inputs are always uncertain.

The Kelly formula, in its simplest form for even-money-style bets, is edge divided by odds: the fraction to bet equals win probability minus loss probability divided by the reward-to-risk ratio. For a strategy winning 50 percent of trades at 2:1 payoff, full Kelly suggests a large fraction of capital per trade, far larger than most traders can tolerate. That is the point and the problem: Kelly maximises long-term growth rate under the assumption that the win rate and payoff are known exactly and never change. In trading they are estimated from a limited sample, they drift as markets change, and a small overestimate of edge produces a large overestimate of size. Full Kelly also produces gut-wrenching drawdowns, with peak-to-trough falls that most humans, and many bots' owners, abandon before recovery.

Fixed fractional risk avoids that fragility by design. Risking a fixed 0.5 to 1 percent of current equity on every trade means position size shrinks automatically during drawdowns and grows slowly during winning streaks, without any estimate of edge beyond the stop distance. A ten-trade losing streak at 1 percent risk leaves about 90.4 percent of the account, needing about 10.6 percent to recover, while the same streak at 10 percent risk leaves about 34.9 percent, needing roughly 187 percent to recover. That recovery arithmetic is why professionals obsess over per-trade risk: losses are easy to take and brutally hard to earn back, and fixed fractional sizing keeps the maths survivable even when the strategy hits a bad patch.

The practical compromise used by many bot operators is fractional Kelly: calculate the Kelly fraction from walk-forward, out-of-sample results, then trade a quarter or a half of it, with a hard cap such as 1 to 2 percent per trade regardless of what the formula says. That preserves Kelly's insight, which is to bet more when the edge is genuinely larger, while respecting the reality that edge estimates are noisy. It also needs a guard: recalculate the Kelly inputs only on a schedule, such as monthly, from a rolling window of recent trades, rather than letting the bot resize itself after every win or loss, which turns sizing into chasing. And never let sizing override the kill switch: a daily loss limit and a maximum drawdown guard must halt trading even when the sizing formula still wants to bet.

Choose the framework you can actually follow through a drawdown, because the best sizing method is the one that keeps you trading a tested strategy instead of abandoning it at the bottom. For most retail bot operators that means fixed fractional risk at 0.5 to 1 percent, with fractional Kelly as an optional refinement once a strategy has hundreds of live trades and a stable edge. Document the sizing rule in the bot's configuration, test it in simulation including losing streaks, and review it quarterly. Sizing does not create an edge, but it decides whether an edge, once found, survives long enough to pay. No sizing method guarantees profit, and leveraged trading can lose more than the initial stake on some venues, so size for the worst streak you have seen, not the average trade you hope for.