What Is the Kelly Criterion?

The Kelly Criterion, developed by John L. Kelly Jr. in 1956, is a formula for sizing bets to maximize the long-run geometric growth rate of your bankroll. It is not about maximizing expected value in a single bet — it is about maximizing the logarithm of wealth over many repeated bets.

The original Kelly formula applies to binary outcomes (win or lose) with fixed odds:

$$f^* = \frac{bp - q}{b}$$

Where:

  • \(f^*\) = fraction of bankroll to wager
  • \(b\) = net odds received (e.g., 2 means you win 2× your bet)
  • \(p\) = probability of winning
  • \(q = 1 - p\) = probability of losing
Key Property: Betting more than Kelly increases variance faster than it increases expected growth — eventually leading to bankroll ruin. Betting less than Kelly is "safe" but sacrifices growth rate. Full Kelly is the mathematically optimal point.

Simple Example: Coin Flip with Edge

Suppose you have a coin that lands heads 55% of the time (p = 0.55) and you are paid 1:1 (b = 1). Kelly says:

$$f^* = \frac{1 \times 0.55 - 0.45}{1} = \frac{0.10}{1} = 10\%$$

You should bet 10% of your bankroll on each flip. Betting more than 10% shrinks your long-run growth rate despite having positive EV.

Kelly Formula for Poker

In poker, outcomes are not binary — you can win or lose many different amounts. The continuous Kelly formula for games with a known win rate and variance is:

$$f^* = \frac{\mu}{\sigma^2}$$

Where:

  • \(\mu\) = expected win rate per hand (or per session) in big blinds or dollars
  • \(\sigma^2\) = variance per hand (or per session) in the same units squared

This gives the fraction of your bankroll to put at risk in a single session. In practice, \(f^*\) tells you the correct game stakes relative to your bankroll.

For example, if your win rate is 5 bb/100 and your variance is 100 bb²/100 hands (standard deviation of 10 bb/100), then:

$$f^* = \frac{5}{100} = 0.05 = 5\%$$

You should play stakes where your total buy-in is at most 5% of your bankroll. At \$1/\$2 NL with a \$200 buy-in, this implies a minimum bankroll of \$4,000.

Three Worked Examples

Example 1: Coin Flip with 3:2 Odds

Setup: You win \$3 for every \$2 wagered (b = 1.5), win probability = 50% (p = 0.5).

$$f^* = \frac{1.5 \times 0.5 - 0.5}{1.5} = \frac{0.75 - 0.5}{1.5} = \frac{0.25}{1.5} = 16.7\%$$

Bet 16.7% of bankroll per wager. With a \$10,000 bankroll, risk \$1,670 per bet.

Example 2: Poker Game with Known Win Rate

Setup: NL100 (\$0.50/\$1.00). Win rate = 8 bb/100 hands = \$8/100 hands. Variance = 2,500 bb²/100 = \$2,500/100 hands.

$$f^* = \frac{8}{2500} = 0.0032 = 0.32\%$$

Only 0.32% of bankroll per 100-hand session should be at risk. With a typical 100-hand session buy-in of \$100 (100bb), the required bankroll is:

$$\text{Bankroll} = \frac{\$100}{0.0032} = \$31,250$$

This seems enormous! This is why full Kelly is impractical for poker — the high variance requires a huge bankroll to be Kelly-optimal. Fractional Kelly is the answer.

Example 3: Tournament Buy-In Decision

Setup: A tournament costs \$500 to enter. Your estimated ROI (return on investment) is 20% (you average \$600 return on a \$500 buy-in). The variance of tournament results is very high — approximated as \$EV × (1/ROI) × k, where k ≈ 5 for standard variance.

Using the simplified Kelly for tournaments:

$$f^* \approx \frac{\text{ROI}}{\text{variance multiplier}} = \frac{0.20}{5} = 4\%$$

The \$500 buy-in should represent at most 4% of your bankroll → minimum bankroll of \$12,500 for this tournament. Most professionals use 1–2% per tournament (half-Kelly or quarter-Kelly), suggesting a \$25,000–\$50,000 bankroll for consistent \$500 tournament play.

Fractional Kelly

Full Kelly maximizes long-run growth but produces terrifying variance — drawdowns of 50% or more are common. Most professionals use fractional Kelly: betting a fixed fraction of the Kelly optimal amount.

$$f_{\text{fractional}} = k \times f^*, \quad k \in (0, 1]$$

Half Kelly (\(k = 0.5\)) produces approximately 75% of the growth rate of full Kelly, but with far less variance. Quarter Kelly (\(k = 0.25\)) is ultra-conservative but virtually eliminates ruin risk.

Kelly FractionGrowth Rate (% of Full Kelly)Max Drawdown RiskRecommended For
Full (1.0×)100%Very High (~50%+)Mathematical ideal only
Half (0.5×)~75%Moderate (~25%)Experienced pros
Quarter (0.25×)~44%Low (~12%)Most recreational players
Tenth (0.1×)~19%Very LowUltra-conservative
Practical Recommendation: Most poker players should use quarter-Kelly to half-Kelly. The growth rate sacrifice is small compared to the psychological and financial benefit of reduced variance and smaller drawdowns.

Bankroll Requirements Table

The following table shows the required bankrolls at various win rates and risk-of-ruin levels for NL cash games, using half-Kelly sizing (buy-in = 100bb).

Win Rate (bb/100)Std Dev (bb/100)5% RoR Bankroll1% RoR Bankroll
280~250 buy-ins~400 buy-ins
580~80 buy-ins~130 buy-ins
880~50 buy-ins~80 buy-ins
1080~40 buy-ins~65 buy-ins
1580~25 buy-ins~40 buy-ins

The risk-of-ruin formula for a given bankroll \(B\) and half-Kelly sizing is approximately:

$$\text{RoR} \approx e^{-2 \mu B / \sigma^2}$$

To find the required bankroll for a 5% RoR: set \(\text{RoR} = 0.05\), solve for \(B\): \(B = \frac{\sigma^2 \ln(20)}{2\mu}\).

Practical Usage

Apply Kelly thinking to three key poker decisions:

  1. Game Selection: Use Kelly to determine which stakes you can afford. If your bankroll supports Kelly-optimal play at NL100, moving to NL200 without doubling your bankroll is overbetting — it increases variance faster than growth.
  2. Shot-Taking: When moving up in stakes, a "shot" of 5–10 buy-ins at the new level is the Kelly-aware approach — large enough to get a sample, small enough to preserve the bankroll if it fails.
  3. Tournament Scheduling: The Kelly approach suggests playing more lower-buyin events and fewer high-buyin events unless your ROI is very high. The high variance of tournaments penalizes over-buyin decisions severely.
Warning: Never estimate your win rate optimistically when applying Kelly. Overestimating your edge leads to over-betting, which increases ruin risk dramatically. Use conservative, well-tracked estimates over at least 100,000 hands.