The Kelly Criterion gives the position fraction that maximizes long-run logarithmic growth of capital given a known win rate and reward/risk ratio.

Formula

kelly_pct = W - (1 - W) / R

W = win rate (probability of winning)
R = average win / average loss

Worked example: 55% win rate, 1.5 R/R.

kelly = 0.55 - (0.45 / 1.5)
      = 0.55 - 0.30
      = 0.25  -> 25% of capital per trade

Shape

"position_sizing": {
  "method": "kelly",
  "estimated_win_rate": 0.55,
  "estimated_rr_ratio": 1.5,
  "max_position_pct": 25
}

Params

estimated win rate - expected win rate (0-1).
estimated R/R ratio - expected average win / average loss.
max position percent - hard cap on position size.

Adaptive estimates

The engine uses actual trade history (last 50 trades) to compute W and R when enough history exists. The estimates above are fallbacks for cold-start.

Why most traders don't run full Kelly

Kelly assumes the inputs are exactly right. Real-world estimates are noisy, and full Kelly under noisy inputs has very high drawdowns. See half_kelly and quarter_kelly for the safer variants.

Pitfalls

Overestimating win rate kills you faster than underestimating. A 5pp overestimate at high R can produce ruin in a few bad trades.
No volatility awareness. Kelly sizes the same regardless of current ATR; consider stacking volatility_adjusted if your stop varies in price terms.