Practice17 May 2026 · 6 min read

Position Sizing Is Not Optional — It Changes Everything

Why position sizing determines whether a positive-expectancy system survives, and the three-layer sizing stack worth implementing.

Two traders run the same signal on the same market over the same decade. One ends with a 40% CAGR and a 22% max drawdown. The other ends with a 6% CAGR and a 71% drawdown, blown out twice along the way. The signal is identical. The difference is position sizing — and most retail systematic traders treat it as a parameter to tune last, after the "real work" of finding edges.

This is backwards. Position sizing is not a refinement layer on top of a strategy. It is the strategy, or at least half of it. A mediocre edge sized correctly outperforms a strong edge sized recklessly, and the math is not subtle.

The expectancy illusion

A positive-expectancy system loses money when sized incorrectly. This is not a paradox — it is a direct consequence of multiplicative returns. Equity compounds geometrically, not arithmetically, so the geometric mean of your per-trade returns is what matters, and the geometric mean is always less than the arithmetic mean whenever variance exists.

G = (1 + r₁)(1 + r₂)...(1 + rₙ) ^ (1/n) - 1

The penalty for variance is structural. A system with a 60% win rate that risks 25% per trade has positive expectancy on paper and a near-certain path to ruin in practice. The arithmetic expectation tells you nothing about whether you survive the journey to collect it.

Volatility scaling: the minimum viable approach

The simplest defensible sizing rule is inverse volatility. Size each position so that its expected dollar volatility contribution to the portfolio is constant. If asset A has 40% annualized vol and asset B has 15%, B gets roughly 2.67x the notional exposure of A to equalize their risk contributions.

position_size = target_risk × portfolio_value / (σ_asset × price)

This single change — moving from fixed-notional to vol-targeted sizing — typically does more for risk-adjusted returns than any signal refinement. It also removes the implicit bet that your highest-vol asset deserves the same exposure as your most stable one, which fixed-notional sizing silently assumes.

If your backtest improves dramatically when you switch from fixed-dollar to vol-targeted sizing, your "edge" was partly a passive bet on the lowest-vol asset in your universe. This is worth knowing before you trade it.

Kelly and why you should not use it

The Kelly criterion gives the position size that maximizes the geometric growth rate of capital. For a system with edge e and variance σ², the Kelly fraction is approximately e/σ². It is mathematically optimal under strong assumptions: known distribution, stationary parameters, no estimation error.

None of those assumptions hold for any real strategy. Your estimated edge is noisy. Your variance estimate is noisy. The distribution has fatter tails than you measured. The result is that full Kelly sizing on estimated parameters produces ruinous overbetting in practice — the literature consistently finds that traders who target Kelly end up well past it.

Half-Kelly is not "conservative Kelly" — it is closer to what Kelly would actually prescribe if it could see your parameter uncertainty. Fractional Kelly (quarter to half) is the operating range for systems with realistically estimated edges. Full Kelly is a theoretical reference point, not a target.

The drawdown geometry you cannot ignore

Drawdowns compound asymmetrically. A 20% loss requires a 25% gain to recover. A 50% loss requires 100%. A 75% loss requires 300%. This is not a psychological observation — it is arithmetic, and it dictates that maximum drawdown is the binding constraint on most sizing decisions, not expected return.

recovery_required = 1 / (1 - drawdown) - 1

This is why position sizing optimization should target the full return distribution, not the mean. A sizing scheme that increases expected CAGR by 3% while doubling the 95th-percentile drawdown is almost always a worse choice, because the drawdown tail is what determines whether you keep running the system at all. Strategies are abandoned during drawdowns, not during quiet periods.

Practical implementation

A workable position sizing stack for a multi-asset systematic strategy has three layers. First, per-position volatility targeting to equalize risk contributions across instruments. Second, a portfolio-level vol target that scales gross exposure up or down based on realized portfolio volatility versus the target. Third, a hard drawdown circuit breaker that cuts exposure when realized drawdown breaches a threshold.

Each of these layers is itself a research question. What lookback for volatility estimation? Exponentially weighted or rolling? What target portfolio vol? What drawdown threshold and what reduction rule? In Kestrel Signal, these are first-class objects in the backtest configuration, not afterthoughts wrapped around a signal — because the joint behavior of signal and sizing is what you are actually testing.

Test your signal at three sizing schemes before trusting any backtest result: fixed-notional, vol-targeted, and vol-targeted with a portfolio cap. If the signal looks good only under one of them, you have not found an edge — you have found a sizing artifact. If it survives all three with reasonable parameter sensitivity, you have something worth trading.