Applying Kelly in High-Vol Crypto: Calibrating Quarter-Kelly in Practice
Full Kelly is almost never used in crypto — why? With 24 months of data across six strategies, we quantify the realised growth and drawdown of 1/2, 1/4, 1/8 Kelly across volatility regimes.
— M. Zhao
The Kelly Criterion's f* = (b·p − q) / b maximises the long-run logarithmic growth rate of wealth — but Kelly's 1956 setup carries two hidden assumptions: the win probability p and the odds b are known constants, and outcomes are i.i.d. Drop those into 2026 crypto markets, where p and b are noisy rolling estimates and visibly non-stationary, and what happens?
The empirical answer: running full Kelly underperforms half- and quarter-Kelly. This is not a hunch; it is a result that the fractional-Kelly literature confirms repeatedly. This essay is a practical calibration for crypto: under what volatility regime, what correlation structure, is f*/4 the right safety margin?
Full Kelly · f*
FULL KELLYGrowth-optimal against the true (p, b). But your estimates (p̂, b̂) carry variance, and Kelly is brutally sensitive to estimation error — a 5% overestimate of p̂ can blow up f* by 40%.
Half Kelly · f*/2
HALF KELLYProvable: under estimation noise, f*/2 loses only ~25% of the expected growth rate but cuts the expected maximum drawdown roughly in half. The institutional quant's typical starting point.
Quarter Kelly · f*/4
QUARTER KELLYOur default sizing cap. In a market with large estimation error and visibly non-stationary volatility — especially leveraged derivatives strategies — quarter-Kelly is our empirically validated anchor.
We backtested the F* Reference Fund's six strategies over a 24-month window: funding-rate arb, cross-market arb, options market-making, event-driven, CTA trend, stablecoin base yield. Each strategy was run at f*, f*/2, f*/4 and f*/8, and we recorded annualised geometric return and maximum drawdown.
Geometric mean is not monotone in sizing — there is a clearly visible interior maximum. In our data, f*/2 reports the highest geo mean (14.4%), but f*/4's ratio of geometric mean per unit of drawdown is materially better than every other bucket. If your investor's risk preference is not just "maximise returns" but "maximise compounding under a drawdown < 20% constraint", f*/4 wins almost every time.
F* Protocol's built-in KellyPolicy uses f*/4 as its default sizing cap; integrating funds may change to f*/2 or a custom ratio via governance, but the change requires multisig + Timelock. Quarter-Kelly is not a recommended parameter — it is a protocol-level engineering constraint that has been made governable.
In a high-vol market with hard-to-estimate parameters, full Kelly is a theoretical limit — what matters is knowing where it sits and leaving enough engineering margin around it. f*/4 is not "conservative"; it is "optimal after accounting for estimation error and non-stationarity". This is the single most important engineering decision F* Protocol makes when translating Kelly's math into on-chain settlement logic.
Backtest code and dataset are open-sourced at fund-vault/research/kelly-quarter/. Results are re-run and archived in every quarterly protocol release.