Home › Brands & organisations › 賽馬數據分析 › Kelly Criterion

Kelly Criterion

The formula that sizes a bet to your edge

The Kelly criterion is a formula for sizing a bet in proportion to the edge you hold. It was introduced in 1956 by John L. Kelly Jr. at Bell Labs, originally as a solution to a problem in information theory, and was later applied to blackjack and markets by Edward Thorp.

[ edit intro ]

🧮 The formula

[ edit ]

f* = (b × p − q) ÷ b, where p is the probability of winning, q = 1 − p, and b is the net odds received (b = 1 means even money). Equivalently, f* = edge ÷ odds.

🎯 Why that fraction

[ edit ]

Kelly maximises the expected logarithm of wealth, which is the long-run growth rate of capital. Bet too little and the edge is wasted; bet too much and a losing streak cripples the bankroll.

📈 Continuous form

[ edit ]

For portfolios the formula becomes f* = μ ÷ σ², which explains why more volatile assets warrant smaller allocations even when expected returns are higher.

⚖️ The cost of overbetting

[ edit ]

Betting twice the Kelly fraction drives the long-run growth rate to zero; betting more than twice makes it negative — even when every bet has positive expected value.

🎲 A famous experiment: the 60% coin

[ edit ]

In 2016 Haghani and Dewey gave 61 young, quantitatively trained participants US$25 and 30 minutes to bet on a coin they were told came up heads 60% of the time. The Kelly-optimal stake is 10% of capital on each flip.

Most participants bet poorly, and about 30% went bust, losing the full stake. Even trained professionals struggle to size bets optimally in the simplest possible game with a clear edge — which is why practitioners use fractional Kelly.

🧩 Fractional Kelly

[ edit ]

In practice bettors use half or quarter Kelly, because p is estimated rather than known and overestimating the edge leads directly to overbetting.

🐎 Horse racing

[ edit ]

William Benter's betting system used Kelly to size bets after a model estimated each runner's probability and compared it with the pari-mutuel odds. In Hong Kong the Jockey Club's 17.5% takeout must be subtracted before any edge is computed.

🔗 Related entries

[ edit ]

❓ FAQ

[ edit ]

Does Kelly guarantee a profit?
No. It maximises growth only if the edge is real; a false edge simply loses money faster.

Why not bet full Kelly?
Because the true probability is unknown; estimation error makes full Kelly dangerous.

📚 References

[ edit ]
  1. Kelly, J. L. (1956). A New Interpretation of Information Rate (PDF)
  2. Thorp, E. O. (2006). The Kelly Criterion in Blackjack, Sports Betting, and the Stock Market (PDF)
  3. Benter, W. (1995). Computer Based Horse Race Handicapping and Wagering Systems (PDF)
  4. Haghani, V. & Dewey, R. (2017). Rational Decision-Making Under Uncertainty: Observed Betting Patterns on a Biased Coin (arXiv)
If you spot an error or have an editorial suggestion, please submit a correction.