What the Kelly Criterion actually solves
Most bettors obsess over which side to take and barely think about how much to stake. That is backwards. Once you have a genuine edge, bet sizing determines almost everything about your long-run results: how fast your bankroll grows, how deep your drawdowns get, and whether a rough stretch ends your season or just dents it. The Kelly criterion is the math that ties stake size to edge. It was published by Bell Labs researcher John L. Kelly Jr. in 1956, and it answers one question precisely: what fraction of your bankroll maximizes long-term growth without ever risking ruin?
The short version: bet more when your edge is bigger, bet less when the odds are longer, and never bet so much that a normal losing streak can wipe you out. The rest of this article makes that concrete with real numbers.
The formula, in plain terms
For a simple win-or-lose bet, the Kelly fraction is:
f* = (b × p − q) / b, which simplifies to f* = p − q/b
- p = your estimated probability the bet wins
- q = 1 − p (probability it loses)
- b = the decimal odds minus 1 (your profit per $1 staked if you win)
An equivalent and more intuitive form is f* = edge / b, where edge is your expected profit per $1 risked. Two inputs, nothing else: how likely you are to win, and how much you get paid. Notice what is not in the formula — how confident you feel, how the last bet went, or how badly you want action.
Worked example 1: a -110 favorite
American odds of -110 convert to decimal 1.909, so b = 0.909. Suppose your model says the true win probability is 55% (p = 0.55, q = 0.45).
Expected value per $1: 0.55 × 0.909 − 0.45 = 0.50 − 0.45 = +0.05, a 5% edge.
Kelly fraction: f* = 0.55 − (0.45 / 0.909) = 0.55 − 0.495 = 0.055.
Full Kelly says stake 5.5% of your bankroll. On a $2,000 bankroll that is $110. Reasonable — but as you will see, most disciplined bettors would stake less than the full figure on purpose.
Worked example 2: a +150 underdog
+150 is decimal 2.5, so b = 1.5. The odds imply a 40% breakeven (100 / 250). Suppose you think the real probability is 45% (p = 0.45, q = 0.55).
Edge: 0.45 × 1.5 − 0.55 = 0.675 − 0.55 = +0.125, a 12.5% edge.
Kelly fraction: f* = 0.45 − (0.55 / 1.5) = 0.45 − 0.367 = 0.083, or 8.3% of bankroll.
Bigger edge, bigger stake — exactly what you would expect.
Why longer odds mean smaller bets at the same edge
Here is the subtlety people miss. Take two bets, each with an identical 5% edge:
- Bet A at -110 (b = 0.909): f* = 0.05 / 0.909 = 5.5%
- Bet B at +200 (b = 2.0): f* = 0.05 / 2.0 = 2.5%
Same edge, but Kelly stakes less than half as much on the longshot. Why? Longshots win less often, so the same expected value comes with far more variance and deeper swings. Kelly automatically discounts for that volatility. Flat-staking every play at "one unit" ignores this entirely and quietly overbets your longshots.
Where does p come from? De-vigging
Kelly is only as good as your probability estimate, and a sportsbook's posted odds are not a fair probability — they include the vig (the book's built-in margin). You have to strip it out first.
Say a two-way market reads Team A -130, Team B +110. Convert to implied probabilities: -130 → 130/230 = 56.52%; +110 → 100/210 = 47.62%. They sum to 104.14% — that extra 4.14% is the hold. Divide each by the total to get the no-vig fair price:
- Team A: 0.5652 / 1.0414 = 54.3%
- Team B: 0.4762 / 1.0414 = 45.7%
Now suppose a different book offers Team A at -110 (52.4% implied). Against a fair 54.3%, you have an edge: EV = 0.543 × 0.909 − 0.457 = +0.036, or 3.6%. Full Kelly = 0.036 / 0.909 ≈ 4% of bankroll; half Kelly, 2%. That full chain — de-vig the market to a fair price, shop for a better number, measure the edge, then size it — is exactly what EdgeFinder automates across 40+ books in MLB, NBA, NFL, NHL, college, and soccer. You can see the current output on today's edges without doing the arithmetic by hand.
Full vs. fractional Kelly
Full Kelly maximizes growth only if your probability estimate is exactly right. In real betting it never is — p is an estimate with error bars. And Kelly punishes overestimation brutally, because the growth curve is a downward parabola. Bet at twice the Kelly fraction and your expected long-run growth rate drops to zero, even with a genuine edge. Bet beyond 2x and a real, provable edge still trends toward going broke. Overbetting doesn't just lower returns; it can flip a winning strategy into a losing bankroll.
That is why most serious bettors use fractional Kelly — typically a quarter to a half of the full figure. The trade-off is favorable: half Kelly captures roughly 75% of full Kelly's growth rate while cutting variance and drawdowns by about half. You give up a little upside to buy a lot of survivability. Given that your p is uncertain, fractional Kelly also builds in a margin of safety against your own estimation error.
Practical bankroll guardrails
- Use fractional Kelly by default — quarter or half. Treat full Kelly as a theoretical ceiling, not a target.
- Cap any single bet at a hard number (say 1–3% of bankroll) regardless of what Kelly suggests. This protects you when a fat "edge" is really a bad estimate.
- Shade your probability toward the market. The closing line is the sharpest public estimate that exists; if your number is wildly different, the market is usually right.
- Resize periodically, not every bet. Recompute your unit off your current bankroll weekly, not after every win or loss — chasing bet-to-bet swings reintroduces the emotion Kelly is meant to remove.
- Treat correlated bets as one exposure. Kelly assumes independent outcomes; parlays and same-game combos are not independent, so size them smaller.
- Track closing-line value (CLV). If you consistently beat the closing number, your edges are probably real and your Kelly inputs are trustworthy. If you don't, your p estimates — and therefore your stakes — are suspect.
That last point is why a verifiable record matters. EdgeFinder logs its model's pregame predictions with timestamps and grades them publicly on win rate, ROI, and CLV, so the probabilities feeding any sizing decision are auditable rather than asserted. You can read the methodology and create a free account to see the top picks; the free tier is a fine place to sanity-check your own numbers before you stake anything.
The honest caveat
Kelly optimizes bankroll growth assuming you truly have an edge. It is not a promise of profit, and no sizing rule turns a losing model into a winning one — it only keeps a winning model from blowing up. Your probability estimates will sometimes be wrong, variance is real, and even correctly sized bets go through losing stretches. Bet only discretionary money you can afford to lose, never funds you need. This is education, not advice, intended for readers 21+. If gambling stops being fun, call 1-800-GAMBLER.