The Kelly Criterion converts a probability edge into a suggested fraction of bankroll. Under strict assumptions—known probabilities, repeatable independent opportunities, no limits or friction—it maximizes expected logarithmic bankroll growth over the long run.
Those assumptions are much stronger than they first appear. In sports betting, probabilities are estimated, markets move, bets correlate, and limits constrain execution. Full Kelly can therefore recommend stakes that are mathematically correct for the input but dangerously large when the input is wrong.
This guide explains the formula, shows practical calculations, and treats fractional Kelly and hard caps as essential risk controls rather than optional details.
What the Kelly Criterion Optimizes
Kelly does not maximize:
- the chance of winning the next bet;
- expected profit on one wager;
- the smoothness of the bankroll;
- protection from every drawdown.
It maximizes the expected logarithm of wealth under a repeated-bet model. The logarithm penalizes ruin heavily and captures the compounding nature of bankroll growth.
Betting too little sacrifices potential growth when an edge is real. Betting too much increases volatility and can reduce long-run growth—even when the wager has positive EV. Full Kelly identifies the theoretical balance for its assumptions.
The Basic Kelly Formula
For a two-outcome wager:
f* = (b × p − q) ÷ b
Where:
f*= full Kelly fraction of current bankroll;b= net decimal odds, or decimal odds minus 1;p= estimated probability of winning;q= estimated probability of losing,1 − p.
An equivalent form is:
f* = (p × Decimal odds − 1) ÷ (Decimal odds − 1)
If f* is zero or negative, Kelly says not to bet. It does not recommend betting the other side unless that side has been modelled separately with positive edge.
Worked Example 1: Odds 2.20, Probability 55%
Given:
- decimal odds: 2.20;
b = 2.20 − 1 = 1.20;p = 0.55;q = 0.45.
Calculation:
f* = (1.20 × 0.55 − 0.45) ÷ 1.20
f* = (0.66 − 0.45) ÷ 1.20 = 0.175
Full Kelly suggests 17.5% of bankroll.
For a 2,000-unit bankroll:
2,000 × 0.175 = 350 units
That is an extremely large real-world sports bet. The number illustrates Kelly’s sensitivity; it should not be read as a general recommendation.
Why the Suggested Stake Is So Large
Odds 2.20 imply a 45.45% break-even probability. An estimate of 55% represents a 9.55-percentage-point gap and theoretical EV of:
(0.55 × 2.20) − 1 = +21%
A genuine, repeatable 21% edge is enormous. Full Kelly treats the estimate as known truth and sizes accordingly. In an efficient sports market, such a large apparent edge should trigger suspicion about the model, data, settlement rules, or price availability before it triggers a large stake.
Worked Example 2: A Smaller Edge
Suppose:
- decimal odds: 1.95;
- estimated probability: 53%;
b = 0.95;p = 0.53;q = 0.47.
f* = (0.95 × 0.53 − 0.47) ÷ 0.95
f* = (0.5035 − 0.47) ÷ 0.95 = 0.0353
Full Kelly is approximately 3.53%.
| Bankroll | Full Kelly | Half Kelly | Quarter Kelly |
|---|---|---|---|
| 500 | 17.63 | 8.82 | 4.41 |
| 1,000 | 35.26 | 17.63 | 8.82 |
| 5,000 | 176.32 | 88.16 | 44.08 |
Even a modest probability disagreement can produce a meaningful stake.
Full, Half and Quarter Kelly
Fractional Kelly multiplies the full recommendation by a chosen fraction.
Fractional stake = Full Kelly fraction × Kelly multiplier
| Method | Multiplier | If full Kelly is 8% |
|---|---|---|
| Full Kelly | 1.00 | 8.0% |
| Half Kelly | 0.50 | 4.0% |
| Quarter Kelly | 0.25 | 2.0% |
| Eighth Kelly | 0.125 | 1.0% |
Fractional Kelly reduces growth in the ideal model but also lowers volatility and the cost of probability error. It does not make a bad estimate safe; it only reduces exposure.
Probability Error Changes Kelly Dramatically
Keep odds fixed at 2.00.
| Estimated win probability | EV | Full Kelly | Half Kelly |
|---|---|---|---|
| 48% | −4% | 0% | 0% |
| 50% | 0% | 0% | 0% |
| 51% | +2% | 2% | 1% |
| 53% | +6% | 6% | 3% |
| 55% | +10% | 10% | 5% |
| 60% | +20% | 20% | 10% |
At even money, full Kelly equals 2p − 1. An error of four probability points can change the suggested stake by eight bankroll percentage points.
If you estimate 55% but the real chance is 50%, a 10% full-Kelly stake is not merely suboptimal—it is being placed with no edge.
Overbetting Kelly Is Especially Costly
Under Kelly’s ideal assumptions, betting more than full Kelly reduces expected logarithmic growth. At sufficiently high multiples, expected growth becomes negative despite a positive-value bet.
For a 55% chance at odds 2.00:
- full Kelly: 10%;
- double Kelly: 20%;
- risking 100%: possible immediate ruin on a loss.
The bankroll multiplier after a win is 1 + f; after a loss it is 1 − f at even money. A 20% stake produces multipliers of 1.20 and 0.80. Large losses require disproportionate recovery, so aggressive overbetting damages compound growth.
Worked Example 3: Sensitivity Range
You can buy odds of 2.10 and estimate 52%, but believe the plausible range is 49%–55%.
| Probability | EV | Full Kelly |
|---|---|---|
| 49% | +2.9% | 2.64% |
| 52% | +9.2% | 8.36% |
| 55% | +15.5% | 14.09% |
The central estimate recommends 8.36%, but the lower plausible case recommends only 2.64%. A conservative process might use quarter Kelly on the central case, cap the stake below that, or pass if the uncertainty cannot be justified.
Sensitivity analysis should be performed before choosing the Kelly fraction.
Estimation Risk Is the Central Problem
Sports probabilities are not known constants. Error comes from:
- small samples;
- changing lineups and roles;
- model misspecification;
- data quality;
- unmodelled correlation;
- market and price latency;
- overfitting;
- subjective adjustments.
Kelly assumes p is correct. When it is estimated, the mathematically optimal stake is uncertain too.
Calibration should be tested on out-of-sample forecasts. If selections labelled 60% win only 54%, full Kelly systematically overbets them.
Kelly Requires the Current Price
Suppose probability remains 53%:
| Decimal odds | EV | Full Kelly |
|---|---|---|
| 1.85 | −1.95% | 0% |
| 1.90 | +0.70% | 0.78% |
| 1.95 | +3.35% | 3.53% |
| 2.00 | +6.00% | 6.00% |
The same selection moves from no bet at 1.85 to a 6% full-Kelly stake at 2.00. Use the accepted price, not an opening price, a screenshot, or an unavailable quote.
Bankroll Definition Matters
Kelly’s denominator should be risk capital genuinely available to the strategy. Do not include:
- essential savings;
- credit;
- bonuses that cannot be withdrawn;
- funds already committed to open bets;
- balances at inaccessible or high-risk operators;
- money needed for another purpose.
If several bets are open simultaneously, calculating each as if the full bankroll remains uncommitted overstates available capital.
Multiple Simultaneous Bets
Naively applying single-bet Kelly to every opportunity can create excessive total exposure.
Example:
- Bet A: 6% full Kelly;
- Bet B: 5%;
- Bet C: 4%;
- Bet D: 5%.
Taken independently, they expose 20% of bankroll. If all depend on the same team-rating error or match condition, portfolio risk can be much larger than the individual formulas imply.
True portfolio Kelly requires a joint return distribution and correlations. In practice, use conservative fractions, event-level caps, sport or model caps, and total open-exposure limits.
Correlated Bets
Examples of positive correlation:
- home win and home −0.5;
- over 2.5 and both teams to score;
- player to score and team to win;
- multiple selections generated by the same league model.
Treating them as independent double-counts the same edge. A single adverse match state can make every position lose together.
If correlation cannot be estimated, reducing or eliminating overlapping bets is more defensible than pretending it is zero.
Pushes, Partial Wins and Multi-Outcome Bets
The basic binary formula assumes only full win or full loss. Asian handicaps, draw-no-bet, dead heats and exchange commission create additional return states.
For general discrete outcomes, Kelly chooses fraction f to maximize:
Expected log growth = Σ pᵢ × log(1 + f × rᵢ)
Where rᵢ is the net return per unit in outcome i.
There may be no simple one-line solution. Use the correct payout distribution rather than forcing a quarter-line bet into a binary formula.
A Practical Capped Fractional-Kelly Process
- Define the exact market and payout states.
- Estimate probability before observing the result.
- Check model calibration and uncertainty.
- Record the current accepted odds.
- Calculate full Kelly.
- Apply a conservative fraction such as half, quarter, or less.
- Apply a hard per-bet cap.
- Apply event, league, model and total-exposure caps.
- Recalculate using bankroll net of committed stakes.
- Round down to permitted stake increments.
- Do not increase because recent bets lost.
This process intentionally produces a stake no larger than the formula suggests.
Worked Example 4: Fraction and Hard Cap
Given:
- bankroll: 1,500 units;
- odds: 2.05;
- estimated probability: 52%;
- Kelly policy: quarter Kelly;
- hard cap: 1% of bankroll.
Full Kelly:
b = 1.05, p = 0.52, q = 0.48
f* = (1.05 × 0.52 − 0.48) ÷ 1.05 = 6.29%
Quarter Kelly:
6.29% × 0.25 = 1.57%
Hard cap reduces the stake to 1%:
1,500 × 1% = 15 units
The policy—not the most aggressive formula—determines the final stake.
When Kelly Should Not Be Used
Avoid Kelly sizing when:
- probabilities are based mainly on intuition;
- the model is not calibrated out of sample;
- the odds are stale or not actually available;
- settlement has multiple unmodelled states;
- positions are strongly correlated;
- the bankroll includes essential money;
- the suggested stake creates discomfort or loss chasing;
- records are incomplete;
- betting limits or account risks dominate execution.
Flat small stakes or no bet are valid alternatives.
Common Kelly Mistakes
Using decimal odds as b
b is net odds: decimal odds minus 1.
Entering percentages incorrectly
55% must be entered as 0.55, not 55.
Ignoring negative output
A negative result means zero stake under this model, not a negative bet size.
Rounding probability upward
Small optimistic changes can materially increase Kelly.
Treating full Kelly as a target
It is an upper theoretical optimum under exact assumptions, not a mandatory real-world stake.
Applying every bet independently
This ignores capital already committed and cross-bet correlation.
Updating bankroll after every unconfirmed result
Use settled, accessible funds and a consistent accounting rule.
Kelly vs Other Staking Methods
| Method | Main input | Strength | Weakness |
|---|---|---|---|
| Flat stake | Predefined unit | Simple and auditable | Does not scale with edge |
| Fixed bankroll percentage | Current bankroll | Automatically scales capital | Ignores price-specific edge |
| Full Kelly | Probability and price | Ideal log-growth optimum | Too sensitive to estimation error |
| Fractional Kelly | Probability, price, multiplier | Lower volatility and error cost | Still depends on model quality |
| Progressive system | Previous result | Simple sequence | No connection to current EV |
Frequently Asked Questions
Does Kelly guarantee bankroll growth?
No. It assumes the probability inputs are correct and opportunities repeat under the model. Real results can produce deep drawdowns, and wrong estimates can cause long-term losses.
Is half Kelly always safe?
No. Half of an overstated recommendation can still be too large. Hard caps and model validation remain necessary.
Why does Kelly sometimes recommend 20% or more?
The input describes an unusually large edge. In sports markets, investigate whether the estimate or price is wrong before accepting such a stake.
Should beginners use Kelly?
Usually not with real stakes. Beginners should first learn probability, keep records, test calibration, and use paper or very small flat stakes.
Can Kelly be used for accumulators?
Only if the joint probability and payout are estimated accurately. Multiplying correlated leg probabilities creates incorrect inputs and dangerous stakes.
Final Thoughts
The Kelly Criterion is a powerful connection between edge, price and bankroll growth. It is also unforgiving of false confidence. Full Kelly assumes you know the probability; sports bettors almost never do.
Use the formula as an upper-bound analytical tool. Test calibration, run sensitivity analysis, apply a conservative fraction, cap each bet and total correlated exposure, and use only discretionary capital. When the estimate is uncertain, the best Kelly adjustment is often a much smaller stake—or zero.
Last updated: July 2026
Published by LineScout Betting Academy
