Kelly Criterion in Practice: Dynamic Staking Without Blowing Up
The Kelly Formula (and Why You Shouldn't Use It Fully)
The Kelly criterion tells you the mathematically optimal fraction of your bankroll to bet: K* = W - (1-W)/R where W is win probability and R is your average win/average loss ratio.
The problem? Kelly assumes you know W and R exactly. In reality, you're estimating them from a finite sample of signals. Full Kelly with estimated parameters will blow you up — the variance is simply too large.
Fractional Kelly: The Practical Answer
Most professional quants use half-Kelly or quarter-Kelly:
- Full Kelly (1.0×): Maximizes log-growth but has enormous drawdown risk. A 10% error in your W estimate can turn a 40% edge into a negative one.
- Half-Kelly (0.5×): Captures 75% of the growth with only 50% of the variance. This is the sweet spot for most retail traders.
- Quarter-Kelly (0.25×): For volatile instruments or when your signal estimate is noisy. Growth drops to 62.5% but drawdowns are halved.
Estimating W and R From Signal Data
Where most people go wrong: they use overall win rate. The correct approach is conditional estimation:
Step 1: Segment by signal score
Signals scoring 80+ behave differently than signals scoring 60. Calculate W and R within each score band, not globally.
Step 2: Apply Bayesian shrinkage
If you only have 30 trades in the 80+ band, your W estimate has huge confidence intervals. Shrink toward the overall rate: W_adjusted = (n×W_sample + N×W_prior) / (n + N) where N is your "pseudo-sample" size (typically 10-20).
Step 3: Cap at the position limit
Never let Kelly suggest more than 5% of portfolio in any single stock, regardless of the math. Diversification is a form of insurance the formula doesn't account for.
How GemStox Applies This
Every signal in GemStox includes a suggested position size calculated using fractional Kelly on the score-conditional win rate, capped at 5%. When volatility spikes, the fraction shrinks further. You never have to do this math yourself — it's baked into the signal.
The Practical Checklist
- Calculate win rate per signal score band, not overall
- Apply Bayesian shrinkage if fewer than 30 trades in a band
- Use half-Kelly as default, quarter-Kelly for high-vol names
- Hard-cap at 5% per position
- Re-estimate W quarterly as your sample grows
Position sizing with built-in math.
Every GemStox signal includes a volatility-adjusted position size — the Kelly fraction calculated for you. Start with a $3 Day Pass.
See Position Sizes in Action →