Key takeaways
- The Kelly formula f = (bp − q) / b converts your estimated win probability and the odds into the bankroll fraction that maximizes long-run growth.
- Full Kelly assumes your probability inputs are exact, so most sharp bettors use quarter or half Kelly to limit drawdown from estimation error.
- Half Kelly captures roughly three-quarters of full Kelly's growth rate at about half the volatility.
- Kelly assumes independent bets, so correlated same-game wagers must be sized as one position or capped per event.
- A negative Kelly output means no edge exists and the bet should be passed.
Kelly criterion betting turns your estimated win probability and the odds into an exact stake: the fraction of bankroll that maximizes long-run growth. This guide walks the full math on a realistic bet, then shows why fractional Kelly is the default for real bankrolls, with a table comparing growth and drawdown at quarter and half Kelly.
What is the Kelly criterion in sports betting?
The Kelly criterion is a bankroll-sizing formula that tells you the stake fraction maximizing long-run logarithmic growth, given your estimated win probability and the odds on offer. John Kelly Jr. derived it at Bell Labs in 1956 as a signal-transmission result, and it was later applied to gambling and investing (Kelly 1956, original paper).
For a bettor, its appeal is honesty: you cannot compute a Kelly stake without stating your own probability and edge. If your model says a +120 line wins 48% of the time, Kelly turns that +EV assessment into a concrete unit size. Full Kelly sizes aggressively and assumes your probabilities are exactly right, which is why most sharp bettors default to a fraction of it. The formula itself is one line, and the next section covers the calculation.
What is the kelly criterion betting formula and how do you calculate it?
The formula is f = (bp − q) / b, where f is the fraction of bankroll to stake, b is the net odds per unit wagered, p is your estimated win probability, and q = 1 − p. If the result is negative, you have no edge and should pass.
Converting American odds to decimal
For minus odds like −150, decimal odds = 1 + (100/150) = 1.667, so b = 0.667. For plus odds like +130, decimal odds = 1 + (130/100) = 2.30, so b = 1.30. In both cases, b is simply decimal odds minus 1.
Defining p, q, and b from your own model
p must come from your own estimated win probability, not the sportsbook's implied probability. If you estimate 55% on a −110 price (implied 52.4%), then b = 0.909, q = 0.45, and f = (0.909 × 0.55 − 0.45) / 0.909 ≈ 0.055, or 5.5% of bankroll. Our betting cheat sheet walks through the conversions. The formula is only as honest as your p, which is why CLV tracking matters for calibration.
How do you use the kelly criterion in betting step by step?
Step 1: Get the best available line
Line-shop across your books first. Kelly output is only as good as the price you feed it, and a half-point or five cents of price swings the stake materially. If the best number on the market is -110 when others post -115, use -110.
Step 2: Strip the vig to find the implied probability
Remove the bookmaker's margin to get the no-vig fair price. For a two-way market at -110/-110, each side's implied probability is 52.38%, and the vig-free fair probability is exactly 50%. On a three-way market, divide each implied probability by the book's total overround so the three sum to 100%.
Step 3: Plug your estimated probability into the formula
Say your model makes the -110 side a 53% winner. Edge = 0.53 − 0.5238 = 0.0062. Full Kelly stake = edge / (1 − decimal odds) = 0.0062 / 0.9091 ≈ 0.68% of bankroll. That tiny fraction is the point: a 0.6-point edge at -110 is real but thin, and Kelly scales you down accordingly.
This is also where most bettors get burned by input error, because a 55% estimate instead of 53% nearly triples the stake. Variance, not just expected growth, is what you are managing, and fractional Kelly exists because full Kelly's drawdowns are brutal relative to its growth advantage; the standard treatment of this tradeoff is Thorp's Kelly criterion paper (The Kelly Criterion in Blackjack, Sports Betting, and the Stock Market). If your probability estimates come from a service rather than your own model, insist on auditable inputs: our raw feed and methodology are documented at https://mvpcapping.com/developers, so you can check the CLV record before trusting anyone's numbers.
Step 4: Convert the output to dollars and units
On a $10,000 bankroll, 0.68% full Kelly is $68; quarter Kelly is $17. In units at 1u = $100, that is 0.17u quarter Kelly. Round to a stake your books accept and log the closing line so you can track CLV against the model's edge over time.
When should you use fractional Kelly instead of full Kelly?
Full Kelly assumes your win probabilities are exact. They never are. If you estimate 55% and the true probability is 52%, full Kelly stakes money on edge that does not exist, and the growth curve punishes that error harder than any other sizing method. Betting full Kelly with a shrunken edge means overbetting, which turns a positive-EV model into a long-term bleed.
How estimation error shrinks your real edge
Your input probability carries model error, sample noise, and stale lines. A no-vig market price of -110 implies roughly 52.4%, so if your model says 55%, your estimated edge is about 2.6 points. If half of that estimate is noise, your real edge is closer to 1.3 points, and full Kelly sizing built on 2.6 points stakes roughly double what the true edge supports. Overbetting Kelly is the one configuration with negative expected growth, which is why quarter or half Kelly is the default for anyone honest about their model's error bars.
Fractional Kelly compared: growth vs drawdown
The table below uses illustrative numbers for a bettor with a 2.6-point average edge, showing how each fraction trades growth rate for variance (figures are illustrative, not measured results).
| Kelly fraction | Stake on a -110 bet (2.6pt edge) | Relative growth rate | Relative volatility |
|---|---|---|---|
| Full Kelly | ~5.2% of bankroll | 100% | 100% |
| Half Kelly | ~2.6% of bankroll | ~75% | ~50% |
| Quarter Kelly | ~1.3% of bankroll | ~44% | ~25% |
Half Kelly captures roughly three-quarters of full Kelly's growth at half the drawdown severity, which is the trade most sharps should take. Quarter Kelly suits smaller bankrolls or models with wide confidence intervals. If you want to audit how a service sizes its own picks before trusting the math, the methodology is laid out on our about page.
What are the risks of the kelly criterion betting strategy?
Garbage-in edge: the biggest failure mode
Kelly output is only as good as your win probability input. A -110 market implies about 51.2% per side; strip the vig with a no-vig conversion (removing the book's margin from both sides' implied probabilities, as outlined in Pinnacle's betting resources) and you have the market's baseline. If your model says 55% when fair is 52%, you don't have a 3.8% edge, you have a 1.8% edge, and full Kelly will over-stake by roughly double. That's how sharp bettors hit losing streaks sized for edges they never had.
Correlated and simultaneous bets
Kelly assumes each bet is independent and sized against your full bankroll. Same-game parlays and correlated props violate that: two legs sharing an outcome can both lose from one result, so stacking full-Kelly stakes on each is double exposure. Simultaneous bets across books have the same problem when they touch the same game. Cap aggregate exposure per event, or size the correlated bundle as one position.
The same discipline applies to data sources. If you're pulling probabilities from a feed, audit the grading history before trusting it; our esports coverage shows auto-graded results you can check against closing lines, because a Kelly stake placed on an unaudited model is just variance with extra steps.
How does a kelly criterion betting calculator work and when should you use one?
A Kelly calculator automates one thing: it takes your estimated win probability, the price you're getting, and your bankroll, then solves f = (bp − q) / b and outputs the stake. Any spreadsheet does the same math, so the tool's value is speed across many simultaneous bets, not accuracy.
Before trusting one, check three things. First, does it handle American odds and negative prices correctly, since a sign error at -150 flips your edge. Second, does it support fractional Kelly natively, because you should never bet full Kelly on estimated probabilities; the criterion assumes your inputs are exact, and they never are (see the original formulation at Wikipedia's Kelly criterion entry). Third, does it let you audit the inputs, or does it hide the probability behind a confidence score you can't inspect.
Manual math is fine when you're placing two or three bets a day from a model you trust. A calculator earns its keep when you're line-shopping across books and regrading dozens of prices after steam.
How does Kelly sizing fit with a model-driven betting approach?
Kelly is only as good as the probability you feed it. Overestimate your win rate by two points and full Kelly turns a thin edge into overbetting that bleeds units through variance. Calibration is the prerequisite, not a nice-to-have.
CLV is your first validation layer. If you consistently beat the close across multiple books, your numbers are at least sharp enough to price markets. If you're losing CLV while winning bets, your model is getting lucky and your Kelly stakes are built on noise.
Graded results are the second layer. Every pick needs an outcome attached, win or lose, so you can compare your estimated probabilities against actual frequency. If you called 55% winners at -110 and you're hitting 48% over a few hundred bets, recalibrate before trusting your edge inputs.
This is why we auto-grade every pick and publish CLV on each one, with no deleted losses. The honest gap: our confidence scores are calibrated on our tracked record, so treat them as a starting estimate for your own Kelly math, not a substitute for your model.
FAQ
Is the Kelly criterion actually profitable for sports betting?
Kelly is profitable only if your win probabilities are accurate. It maximizes long-run growth given a real edge, but it does not create one. Overestimated probabilities make full Kelly overbet, which produces negative expected growth. Bettors with calibrated models and a genuine edge use fractional Kelly to capture most of the growth while limiting drawdown from estimation error.
What fraction of Kelly do professional bettors use?
Quarter Kelly and half Kelly are the standard choices. Half Kelly captures roughly three-quarters of full Kelly's growth rate at about half the volatility, while quarter Kelly suits smaller bankrolls or models with wide confidence intervals. Full Kelly assumes your probability inputs are exact, which they never are, so betting it on estimated probabilities risks negative expected growth.
Does Kelly work for parlays and correlated bets?
Not directly. The formula assumes each bet is independent and sized against your full bankroll. Same-game parlays and correlated props violate that: two legs sharing an outcome can both lose from one result, so stacking full-Kelly stakes on each doubles your exposure. Size the correlated bundle as one position, or cap aggregate exposure per event.
How do you size Kelly bets across multiple sportsbooks?
Feed each bet the best available price from your line-shopping, then size it against your total bankroll, not the balance at that book. If two bets touch the same game across books, treat them as correlated and cap combined exposure. Log the closing line on every stake so you can check your estimated edge against CLV over time.