Kelly Criterion in Sports Betting: Optimal Stake Sizing for Professional Bettors
Key Takeaways
- Kelly tells you the exact fraction of your bankroll to bet to maximise long-run growth
- Formula: f* = (bp − q) / b where b = net odds, p = win probability, q = loss probability
- Full Kelly is mathematically optimal but practically dangerous — most professionals use ¼ to ½ Kelly
- Kelly assumes your probability estimates are perfectly calibrated — overestimate your edge and Kelly over-bets, risking ruin
- With a 5% edge on a 1.90 bet, Full Kelly recommends ~5.3% of bankroll — ¼ Kelly = ~1.3%
The Kelly Criterion, developed by physicist John L. Kelly Jr. in 1956, solves a specific problem: given a +EV bet, how much of your bankroll should you stake? Bet too little and you under-grow. Bet too much and variance risks wiping you out.
Kelly's formula finds the mathematically optimal fraction. Used correctly — especially in fractional form — it is the foundation of professional bankroll management.
The Kelly Formula
f* = (bp − q) / b
Where:
- f* = fraction of bankroll to bet
- b = net decimal odds − 1 (i.e., profit per unit staked)
- p = estimated probability of winning
- q = probability of losing (= 1 − p)
Alternative form that is often easier to apply:
f* = (p × d − 1) / (d − 1)
Where d = decimal odds (including stake return)
This second form directly uses decimal odds as displayed by bookmakers.
Worked Example: AH Bet with a 5% Edge
Setup:
- Bet: Arsenal AH −0.5, decimal odds = 1.90
- Your estimated win probability: p = 0.58 (58%)
- Net odds: b = 1.90 − 1 = 0.90
- Loss probability: q = 1 − 0.58 = 0.42
Kelly calculation:
f* = (0.90 × 0.58 − 0.42) / 0.90 = (0.522 − 0.42) / 0.90 = 0.102 / 0.90 = 0.1133 (11.33%)
Full Kelly recommends staking 11.33% of your bankroll on this single bet.
Implied EV check: The market's implied probability at 1.90 = 1/1.90 = 52.6%. Your estimate is 58%. Edge = 58% − 52.6% = 5.4%.
| Kelly Fraction | % of Bankroll | Stake on €10,000 Bankroll |
|---|---|---|
| Full Kelly (1×) | 11.33% | €1,133 |
| Half Kelly (½×) | 5.67% | €567 |
| Quarter Kelly (¼×) | 2.83% | €283 |
| Tenth Kelly (1/10×) | 1.13% | €113 |
Why Full Kelly Is Dangerous in Practice
Kelly is derived under specific assumptions that don't perfectly match real-world betting:
1. Probability Estimation Error
Kelly assumes your probability estimate is perfectly accurate. But betting models carry estimation error — your 58% might be 54% or 62%. Kelly amplifies errors: if your true edge is 2% but you estimated 5%, Kelly will over-bet massively. On a losing run driven by overestimation, the drawdown is severe.
2. No Simultaneous Bets
Kelly's original derivation assumes one bet at a time. In practice, you're often placing 5–20 simultaneous bets on different markets. The full multi-bet Kelly extension requires a correlation matrix of outcomes — computationally complex and impractical for most bettors.
3. Log Utility Assumption
Kelly maximises the logarithm of bankroll growth, not the bankroll itself. This implies infinite risk aversion for ruin. Most professional bettors have income outside betting — their personal utility function differs from pure log-bankroll maximisation.
The Solution: Fractional Kelly
Academic literature and professional practice have converged on using a fraction of Kelly — typically ¼ to ½ — as the optimal real-world approach. Fractional Kelly:
- Reduces variance significantly (¼ Kelly → ~⅛ of Full Kelly variance)
- Preserves most of the growth advantage over flat staking
- Provides buffer against probability estimation error
Fractional Kelly Stake Calculator: Common Scenarios
For Asian handicap bets at typical odds (1.85–1.95), with various estimated edges:
| Your Edge | Odds | Full Kelly % | ½ Kelly % | ¼ Kelly % |
|---|---|---|---|---|
| 2% | 1.90 | 4.4% | 2.2% | 1.1% |
| 3% | 1.90 | 6.7% | 3.3% | 1.7% |
| 5% | 1.90 | 11.1% | 5.6% | 2.8% |
| 5% | 2.10 | 9.5% | 4.8% | 2.4% |
| 8% | 1.90 | 17.8% | 8.9% | 4.4% |
Most systematic bettors operate in the 1–3% stake range per bet relative to bankroll, which corresponds to ¼–½ Kelly on typical edges of 3–5%.
Kelly and Bankroll Compounding
Kelly's key insight is that it treats the bankroll as a dynamic entity. After each bet, your Kelly stake is recalculated based on your new bankroll size. This compounding is what produces optimal geometric growth:
- Win a bet → bankroll grows → next Kelly bet is larger in absolute terms
- Lose a bet → bankroll shrinks → next Kelly bet is smaller
Flat staking (fixed unit size regardless of bankroll) ignores this compounding. Over large samples, Kelly-based staking significantly outperforms flat staking for bettors with genuine edges — at the cost of higher variance.
Practical implementation: recalculate your Kelly stake at the start of each week or month based on current bankroll. Daily recalculation introduces noise and over-trades on variance.
Kelly Criterion with Asian Handicap Push Lines
Asian handicap full-ball lines (AH −1, −2) and quarter-ball lines introduce a third outcome: push. The standard Kelly formula assumes two outcomes. For push lines, a three-outcome Kelly extension is needed:
f* ≈ (p_win × d − 1) / (d − 1 + p_push × d)
Where p_push = probability of the push outcome
In practice, for quarter-ball lines, the push probability on a single leg is typically 10–25% (probability of a draw on AH 0 component, or an exact margin hit on full-ball component). The adjustment from standard Kelly is modest — usually less than 15% difference in recommended stake.
For most bettors: using standard Kelly on Asian handicap is a reasonable simplification. The error from ignoring push outcomes is small relative to the error from probability estimation uncertainty.
Kelly in Professional Betting Practice
The professional approach most commonly seen in Asian handicap betting:
- Estimate true probability for each bet using your model
- Calculate implied probability from the available odds (remove bookmaker margin)
- Calculate edge = your probability − implied probability
- Apply ¼ Kelly to the edge and current bankroll to get stake
- Cap maximum bet at 3–5% of bankroll regardless of Kelly calculation (risk management override)
- Track actual P&L vs EV over 500+ bet samples to verify calibration
Access to sharp Asian books is essential for this approach. PS3838 and Pinnacle prices provide the most accurate market probability references. Both are accessible via a betting broker.
Frequently Asked Questions
What if Kelly recommends 0% or negative?
If the Kelly formula returns 0% or negative, the bet is not +EV by your probability estimate. Don't place it. Negative Kelly literally tells you to lay the bet (bet against it), which requires a betting exchange.
Should I use fixed percentage or Kelly?
Fixed percentage (e.g., always 2% of bankroll) is simpler to implement and avoids the risk of overestimating your edge. Kelly optimises growth if your estimates are accurate. Most professionals start with fixed percentage and transition to fractional Kelly once their model is calibrated over 1,000+ bets.
Can Kelly be used for accumulator bets?
Technically yes, but accumulator bets involve correlated outcomes and compounded probability uncertainty. Most serious bettors avoid accumulators entirely — they represent negative EV in almost all cases due to bookmaker margin applied per leg.
What's the relationship between Kelly and CLV?
Kelly determines how much to bet; CLV helps verify whether your model is generating genuine edges. High CLV over large samples confirms your probability estimates are beating the market — which is the input Kelly needs.