The Kelly Criterion is the most-quoted, most-misused formula in sports betting. It promises to tell you the mathematically optimal stake for any bet given your edge — and it does, under conditions that rarely hold in real life. Used straight out of the textbook, full Kelly will gut your bankroll. Used carefully, with realistic edge estimates and a fractional adjustment, it becomes the cleanest sizing framework available. This guide walks through the formula, the failure modes, and how serious bettors actually deploy it.
What the Kelly Criterion actually solves
Kelly answers one specific question: given a known edge and known odds, what fraction of your bankroll maximizes long-run growth? Not maximum profit on a single bet — maximum compounding rate over many bets. Bet too small and you leave growth on the table. Bet too large and even an edge gets eaten by variance.
The classical formula for a binary bet at decimal odds is:
f* = (bp - q) / b
Where:
- f* is the fraction of your bankroll to wager
- b is the decimal odds minus 1 (i.e., the profit per unit risked)
- p is your estimated probability of winning
- q is 1 - p
A worked example
You think the Bengals should beat the Ravens at home 55 percent of the time with Joe Burrow active. Cincinnati is priced at +110 (decimal 2.10). Plug in:
- b = 1.10
- p = 0.55
- q = 0.45
f* = (1.10 × 0.55 - 0.45) / 1.10 = (0.605 - 0.45) / 1.10 = 0.155 / 1.10 = 0.141
Full Kelly says bet 14.1 percent of your bankroll. On a $5,000 bankroll, that is $705 on a single NFL game. If your true probability is actually 53 percent instead of 55, your "optimal" stake is now far too aggressive — and the variance can crush you long before the edge plays out.
Why full Kelly fails in practice
The Kelly formula assumes you know p exactly. In sports betting, you never do. You estimate it. The estimate has noise. And Kelly's variance is brutal: betting full Kelly, you can expect drawdowns of 50 percent or more even with a real edge. Most bettors cannot stomach that, and they tilt out of their plan long before the math vindicates them.
There are three structural problems with full Kelly:
- Overestimated edge. Bettors are systematically optimistic about their win rate. Kelly compounds the optimism into oversized bets, especially when the pick involves a star name like Mahomes, Lamar Jackson, or Jalen Hurts.
- Correlated bets. Kelly assumes independent wagers. Sunday NFL slates are not independent — weather, injuries, and correlated game flow link outcomes.
- Drawdown psychology. Even with a perfect edge, full Kelly sees regular 40 to 50 percent drawdowns. Almost nobody bets through that without abandoning the strategy.
Fractional Kelly: the practitioner's version
Serious bettors use fractional Kelly — typically a quarter or a half of the formula's recommendation. Half Kelly cuts your expected growth rate by about 25 percent but cuts your variance dramatically. Quarter Kelly is even safer, with most of the upside preserved over a long enough horizon.
Want to test these numbers live? The Kelly criterion guide ships with a working calculator — drag the win-prob and Kelly-fraction sliders to watch a 200-simulation bankroll fan redraw in real time.
Using the Bengals example above, full Kelly said 14.1 percent. The fractional versions:
- Half Kelly: 7.0 percent → $352 on a $5,000 bankroll
- Quarter Kelly: 3.5 percent → $176 on a $5,000 bankroll
That $176 stake is still aggressive by traditional bankroll-management standards (1 to 2 percent flat), but it reflects the fact that you have a real, measurable 5-point edge. The point of Kelly is that bigger edges get bigger stakes — but only proportionally, and only after you've haircut for uncertainty.
Estimating your edge honestly
The whole framework breaks if your p is wrong. Two practical rules:
- Use a model, not a feeling. A model you build in the lab outputs calibrated probabilities you can plug into the formula. Your gut does not.
- Verify calibration before trusting Kelly. If your model says 60 percent, it should win 60 percent of those bets in a backtest. If it predicts 60 and wins 53, your edge is overstated and Kelly will hurt you.
You can check a model's historical accuracy directly on the model leaderboards before deploying real money behind it.
Negative-edge bets: don't do them
If the formula spits out a negative number, the bet has negative expected value and Kelly says stake zero. This sounds obvious, but it is the most-violated rule in betting. Public favorites, primetime overs, and lottery-style parlays are usually negative-EV. A Cowboys -7 ticket is not automatically worth a stake because Dallas is popular; it needs positive expected value after juice. Kelly's first commandment: do not bet what does not have edge.
Multi-bet portfolios
Kelly was originally written for a single bet at a time. Real bettors place several bets per day, sometimes correlated. The fix: when you have multiple simultaneous edges, scale each Kelly stake down proportionally so the total at-risk fraction is still inside your comfort zone (often capped at 10 to 15 percent total bankroll exposed at any one moment). For props that share players or correlated game outcomes — Burrow passing yards plus Ja'Marr Chase receiving yards, for example — treat them as a single bet, not independent ones. We unpack that more in our parlay correlation guide.
Edge cases the textbook formula does not cover
The clean f* = (bp − q) / b only handles binary outcomes at fixed odds. Real betting is messier:
- Push potential. Spread bets at integer numbers (NFL -3) push some fraction of the time. The Kelly stake should weight wins, losses, and pushes separately. Treating pushes as losses understates expected return; treating them as wins overstates it. Use the realized push rate at the relevant key number from a calibrated dataset.
- Limit-affected stakes. Sportsbooks rarely take the full Kelly stake on a sharp bettor. If your book caps you at $200 per game, the formula's $700 recommendation translates to whatever you can actually get down. Real Kelly across a season is min(theoretical Kelly, book limit, your cap).
- Vig sensitivity. Kelly assumes you have already devigged the price. Plugging the raw -110 line into the formula instead of the devigged probability inflates the edge by half the hold. Always devig before computing f*.
- Asymmetric payout markets. DFS contests, prediction markets with non-binary outcomes, and parimutuel pools all have variable payouts. Kelly generalizes to fractional Kelly on logarithmic utility, but the closed form changes — practitioners typically simulate growth instead of using a closed-form.
What full Kelly looks like in practice (a war story)
Imagine a bettor with a $10,000 bankroll and a model showing a 5-point edge on a Sunday slate of seven NFL games. Full Kelly says stake roughly 8 to 12 percent on each game. Naively staking 10 percent on seven correlated games means 70 percent of bankroll exposed in a single afternoon. One bad weather forecast, one shared closing trend that hits the wrong way, and a $10,000 roll is a $4,000 roll by Monday morning.
The fixes are well known but require discipline. Cap total Sunday exposure at 15 percent. Scale each individual Kelly stake by the correlation factor (you can estimate this with a running multivariate model of game outcomes by week). And never bet correlated games as if they were independent draws from independent populations.
Why disciplined bettors still use Kelly
Despite all the failure modes, Kelly remains the right framework. Flat betting wastes a real edge by sizing the same on a 1-point edge as on a 6-point edge. Full Kelly is reckless. The middle path — calibrated probabilities, fractional Kelly, hard cap, devigged inputs, correlation-aware portfolio scaling — captures the math advantage while surviving the variance.
You can prototype the whole pipeline on the model workshop: build the probability model, devig the live line, compute fractional Kelly, log the recommended stake, then compare against the closing line. The end-to-end loop is what makes Kelly safe to deploy. Skipping any step puts you back in textbook-Kelly land.
How to actually deploy Kelly
- Pick a sport and a model with documented win-rate calibration.
- Compute f* for each bet using model probability and current odds.
- Apply a fractional multiplier — start at quarter Kelly, move to half only after several hundred bets confirm your edge.
- Cap at 5 percent of bankroll on any single play, no matter what the formula says.
- Re-baseline your bankroll weekly. Kelly stakes scale with the new number, not the original one.
Common Kelly mistakes
Treating implied odds as your edge
If the book lists Cowboys -3.5 at -110, plugging in p = 0.524 (the implied probability) and using Kelly to compute a stake will always say bet zero — there is no edge against the book's own price. You need your own probability estimate, not the implied one. Kelly is meaningless without an independent model that disagrees with the book.
Forgetting to update bankroll after each bet
Kelly stakes scale with bankroll. If you compute a 3% Kelly stake on $5,000 ($150) and your bankroll grows to $7,000, the next bet at 3% Kelly should be $210 — not $150. Conversely, if you drop to $4,000, the bet should be $120. Most bettors lock their stake size and never recalibrate. That works for flat sizing but defeats the entire point of Kelly, which is to compound at the optimal rate. Update at least weekly, ideally per bet.
Ignoring book limits
Sharp books like Pinnacle and Circa take large bets but rarely move limits up for winners. Soft books (DraftKings, FanDuel) limit winners within weeks of consistent CLV-positive play. Your Kelly stake on a sharp book may be capped by your account history rather than by the formula. Plan for stake compression: if your quarter Kelly is $300 but the book will only take $100, you have to either size at the cap or move volume to a deeper-pocketed venue. The bettor desk tracks book-by-book limit history so you can see which books actually let you scale.
Bottom line
The Kelly Criterion is the right framework for thinking about bet size, but full Kelly is almost never the right number. Use a model to estimate probability, plug it into f* = (bp - q) / b, then take a quarter or half of that and cap at 5 percent. The framework rewards real edge with bigger bets and punishes guesswork with brutal variance — which is exactly the behavior you want.
You can build, calibrate, and stake models against live lines on our NFL picks and other sport pages, with edge versus closing line tracked automatically. If full Kelly still feels too aggressive, pair this with basic bankroll management, stay flat until your sample proves otherwise, and verify the edge holds via the CLV explainer before scaling stakes.
Bet responsibly — set limits, never chase losses.
Price examples and pass rules
Use names as evidence, not decoration. The useful SEO win is that Lamar Jackson, Jalen Hurts, Joe Burrow, Ja'Marr Chase and Josh Allen and Ravens, Bengals, Cowboys, Chiefs and Bills appear inside decisions, thresholds, and internal links instead of being dumped into a keyword list.
- Spread example: if Chiefs-Broncos opens Chiefs -3.5 and your fair number is -2.8, +3.5 is the bet, +3 is a pass, and the moneyline needs roughly +155 or better before it replaces the spread.
- Total example: if a Bills outdoor total opens 46.5 and wind moves from 8 mph to 21 mph, an under projection at 42.8 still needs a playable number; under 45 or better is different from chasing 43.5.
- Futures example: Bengals AFC North +280 is 26.3% before hold. If your fair number is 30%, stake modestly, track portfolio correlation, and avoid stacking every Burrow, Chase, and Higgins bet into the same thesis.
- CLV rule: a good write-up is not enough. Track whether the spread, total, prop, or futures price closed better than your entry before grading the process.
Use closing-line value guide, vig and hold guide, bet tracking workflow to keep the examples attached to measurable prices.
Research note board
Use this table to turn the guide into a decision note. The point is to know when the idea is actionable and when it is only context.
| Angle | Input to verify | Example application | Pass when |
|---|---|---|---|
| Market price | Spread, total, moneyline, prop price, or futures hold | Ravens and Bengals compared through CLV | The price has moved past the number that created the edge |
| Football or sport context | Role, pace, weather, injury status, opponent style | Lamar Jackson role news mapped to the relevant market | The original input changes or remains unconfirmed |
| Review loop | Entry, close, result, and reason code | vig logged with a clear thesis | You cannot explain whether the process beat the market |
Expected bankroll growth at 55% edge
Expected geometric growth of a $100 bankroll under different Kelly multipliers across 1000 bets at p=0.55, decimal=2. Full Kelly maximises long-run growth but produces the deepest drawdowns; fractional Kelly trades growth for variance.
Drawdown by Kelly fraction
Median and 95th-percentile max drawdown by Kelly fraction over a 1000-bet horizon. Halving Kelly almost halves drawdown; quartering it cuts drawdown by ~70%. Figures are illustrative ballparks from the Kelly literature.
