PerpLogBlog
Open App
strategy7 min

Kelly Criterion for Crypto Trading: A Practical Approach

By·Founder & Trader

·

The Kelly Criterion is a formula for optimal bet sizing developed by John Kelly at Bell Labs in 1956. It tells you the fraction of capital to risk on each trade to maximize long-term growth. But applied naively to crypto trading, it will blow your account.

The Basic Formula

Kelly% = W - (1 - W) / R

W = win rate (e.g. 0.55 for 55%)
R = avg win / avg loss ratio

A trader with 55% win rate and 1.5:1 average reward-to-risk gets:

Kelly% = 0.55 - (0.45 / 1.5) = 0.55 - 0.30 = 25%

Full Kelly says risk 25% per trade. In practice, this produces catastrophic drawdowns because the formula assumes perfectly known probabilities. Real trading has uncertainty — your 55% win rate might actually be 48% when the sample is small.

Why Full Kelly Is Dangerous

  • Small sample bias — 50 trades is not enough data to know your true win rate. Kelly is extremely sensitive to estimation errors.
  • Non-stationary markets — Your edge varies across market regimes. A 60% win rate in trending markets might drop to 40% in ranging conditions.
  • Fat tails — Crypto drawdowns are not normally distributed. A -50% day happens more often than models predict.
  • Psychological reality — A 40% drawdown (which full Kelly accepts) will cause most traders to tilt and deviate from their system.

Quarter-Kelly: The Professional Standard

Most quantitative funds use a fraction of Kelly — typically between 1/4 and 1/3. Quarter-Kelly captures 75% of the growth rate of full Kelly while reducing maximum drawdown by roughly half. The tradeoff is slower growth for dramatically better survivability.

With the same 55% / 1.5R stats: Quarter-Kelly = 25% / 4 = 6.25% risk per trade. Still aggressive for most traders, but in the realm of reasonable.

Bayesian Shrinkage for Small Samples

When you only have 20-30 trades in a particular setup or market session, your estimated win rate is unreliable. Bayesian shrinkage addresses this by blending your setup-specific stats with your overall (global) trading stats:

Adjusted Rate = (n × setup_rate + κ × global_rate) / (n + κ)

n = number of trades in this setup
κ = shrinkage strength (typically 10-20)

With only 5 trades in a new setup, the adjusted rate stays close to your global average. After 50+ trades, it converges to the setup-specific rate. This prevents over-sizing on setups where you have limited data.

Segmented Kelly: Context Matters

Your win rate varies by context: market session (Asia vs. NY), asset (BTC vs. altcoins), day of week, and trading setup. A single Kelly fraction for all trades leaves money on the table (under-sizing your best setups) while over-sizing your worst.

Segmented Kelly computes separate fractions for each context:

  • BTC breakout during NY session: 62% WR, 1.8R → size up
  • Altcoin counter-trend during Asia: 44% WR, 1.2R → size down (or skip)

Putting It Into Practice

Computing Kelly fractions manually for every context is impractical. You need an engine that tracks your segmented stats and adjusts sizing recommendations in real time. The PerpLog Adaptive Sizing Engine does exactly this — it computes quarter-Kelly with Bayesian shrinkage across your trading sessions, playbooks, and market conditions, with drawdown constraints and regime detection built in.

PerpLog computes adaptive Kelly sizing across all your trading contexts — sessions, playbooks, and market conditions — with Monte Carlo risk simulation built in.

Try Adaptive Sizing Free

Frequently asked questions

What is a safe Kelly fraction for crypto trading?

Quarter-Kelly (0.25× full Kelly) is the professional standard for discretionary traders. It captures roughly 75% of the long-run growth rate of full Kelly while halving expected drawdown. Many institutional systematic traders use even lower fractions (1/8 to 1/10) when their inputs are noisy.

Why is full Kelly dangerous in real trading?

Full Kelly assumes win rate and reward-to-risk are perfectly known. Real trading inputs are noisy: a 55% win rate measured over 50 trades has a wide confidence interval. Acting on full Kelly with estimation error produces drawdowns that are 40-60% deep, which most traders psychologically cannot endure.

How does Bayesian shrinkage help with small samples?

Bayesian shrinkage blends a setup-specific win rate toward a global prior when the sample is small. With only 20 trades in a new setup, the adjusted rate stays close to your overall average; after 50+ trades, it converges to the setup-specific rate. This prevents over-sizing on noisy data.

Should I use the same Kelly fraction for every trade?

No. A single Kelly fraction across all trades leaves money on the table by under-sizing your best setups while over-sizing your worst. Segmented Kelly computes separate fractions per context (session, asset, day-of-week, setup) so you scale up where you have edge and scale down where you don't.

Related reading