The Kelly criterion for bettors (and why you should bet a fraction of it)

There is a formula that answers the question every bettor asks and almost none answer correctly: how much should I bet? It was published in 1956 by a Bell Labs scientist named John Kelly, and it is the mathematically optimal way to grow a bankroll.

It is also a loaded gun. Used straight, it will put you through drawdowns that make grown adults quit. So the pros use a defanged version. Let me show you both, and why the fraction matters more than the formula.

The lie: size bets by gut

Here is the belief we are breaking:

Bet more on the ones you like, less on the ones you don't, by feel.

Everyone does this. It feels like skill. It is actually noise, because your feeling of confidence is a terrible estimate of true probability, and it peaks exactly when you are most emotionally invested and least objective. Sizing by gut means betting biggest on your favorite team, your revenge spot, the game you cannot stop thinking about, none of which correlates with edge.

Kelly replaces the gut with two numbers: your edge and the odds. If you can estimate those honestly, Kelly tells you the exact fraction of your bankroll that grows your money fastest over the long run. No feelings involved.

The formula, in plain words

The Kelly formula for a bet is:

f = (bp - q) / b

Where:

• f is the fraction of your bankroll to bet.
• b is the odds received, in decimal-minus-one terms (a bet at +150 has b = 1.5, because you win 1.5 times your stake).
• p is your estimated probability of winning.
• q is the probability of losing, which is just 1 minus p.

In words: Kelly bets more when your edge is bigger and when the odds pay more, and it bets nothing when you have no edge. If bp is less than q, the formula goes negative, which is Kelly's way of saying do not bet this, it is negative expected value. That built-in "bet zero on bad bets" is half of why the formula is so respected.

A worked example

You are looking at a bet at +150, so b = 1.5. You honestly believe the true win probability is 45%, so p = 0.45 and q = 0.55.

• bp = 1.5 x 0.45 = 0.675
• bp minus q = 0.675 minus 0.55 = 0.125
• f = 0.125 / 1.5 = 0.083

Kelly says bet 8.3% of your bankroll. On a 1,000 dollar bankroll, that is 83 dollars.

And here is the first alarm bell. 8.3% is a huge bet by any sane bankroll standard, four or more times the 1% to 2% units we recommend in bankroll management. Full Kelly is aggressive, and that is before we talk about the fact that your 45% estimate is probably wrong.

The number: full Kelly's brutal swings

Full Kelly maximizes long-run growth, which sounds like exactly what you want. The problem is what it does to you on the way there. Betting full Kelly produces enormous volatility. It is mathematically common under full Kelly to see your bankroll cut in half at some point, a 50% drawdown, even when your edge is real and everything is going "right" in the long run.

Most humans cannot bet through a 50% drawdown without panicking, chasing, or quitting. And a drawdown that makes you abandon the strategy turns the optimal formula into a losing one, because you bailed at the bottom. Full Kelly is optimal only for a perfectly disciplined bettor with a perfectly accurate edge estimate. You are neither, and neither is anyone.

Why you bet a fraction of Kelly

This is the part that matters more than the formula itself. Serious bettors do not bet full Kelly. They bet half Kelly or quarter Kelly, a fixed fraction of what the formula says.

The trade is spectacular. Half Kelly captures about three-quarters of full Kelly's long-run growth while roughly halving the volatility. You give up a little growth and buy a lot of sleep. Quarter Kelly gives up more growth for even smoother equity. Here is the shape of the tradeoff:

Approach	Long-run growth	Volatility / drawdowns	Who it suits
Full Kelly	Maximum	Severe (50% drawdowns common)	Almost nobody in practice
Half Kelly	~75% of full	Roughly half of full	Disciplined bettors with decent estimates
Quarter Kelly	Lower still	Much smoother	Most real bettors, uncertain estimates

There is a second, deeper reason to bet a fraction. Kelly assumes your probability estimate is exactly right. It never is. Overestimating your edge makes Kelly tell you to overbet, which is far more dangerous than underbetting, because overbetting can push you past the point of no return while underbetting just grows a little slower. Betting a fraction of Kelly builds in a margin of safety against your own overconfidence. Since every bettor overrates their edge, everyone should shade down.

A worked example, fractional

Back to our +150 bet where full Kelly said 8.3%.

• Half Kelly: 4.15% of bankroll, about 42 dollars on 1,000.
• Quarter Kelly: 2.08%, about 21 dollars on 1,000.

Look at quarter Kelly: about 2%, right back in the sane unit range from bankroll management. That is not a coincidence. Quarter Kelly on realistic edges lands close to the 1% to 2% flat unit that experienced bettors use, which is why flat betting small units is a reasonable approximation of fractional Kelly for people who cannot estimate their edge precisely. If you do not trust your probability estimates, flat small units and quarter Kelly are cousins, and both keep you alive.

Where Kelly breaks down

Honest caveats, because Kelly is powerful and over-trusted.

• Garbage in, garbage out. Kelly is only as good as your probability estimate, and yours is uncertain. A confident-but-wrong p leads directly to ruinous overbetting. This is the single biggest failure mode.
• It assumes one bet at a time. Real slates have many simultaneous, sometimes correlated bets. Summing full-Kelly stakes across correlated bets overbets badly. Scale down when you have multiple live positions.
• It ignores your psychology. The math does not know you will panic at a 40% drawdown. If a stake size will make you deviate from the plan, it is too big, whatever Kelly says.
• It assumes a fixed edge. Edges decay as markets move. A stale edge estimate makes Kelly overbet a bet that is no longer there.

Kelly across a full slate of bets

The clean formula assumes one bet at a time, but real betting is a slate: several bets live at once, some on the same games. This is where naive Kelly quietly blows people up, and it is worth understanding before you apply it.

If you compute full Kelly for each of five simultaneous bets and place all of them, you have effectively overbet, because the formula sized each one as if it were your only exposure. Five bets at 8% each is 40% of your bankroll in play at once, which is nowhere near what single-bet Kelly intended. The swings stack. Worse, if any of those bets are correlated, say two legs tied to the same game going the same direction, they can lose together, and your true risk is far higher than five independent positions would suggest. Correlation is the same force that makes the same-game parlay correlation tax so punishing, and it hurts your sizing the same way.

The practical fixes are simple. Cap your total exposure across the slate, treat correlated bets as closer to a single position rather than several, and scale your per-bet fraction down as the number of simultaneous bets rises. A bettor with ten live plays should be sizing each far smaller than the single-bet formula suggests. When in doubt, the safer error is smaller, because underbetting costs you a little growth while overbetting can cost you the bankroll.

There is another practical fix that sounds unglamorous but works: set a maximum unit ceiling before the slate starts. If your usual unit is 1%, maybe no single bet can exceed 1.5% until your tracked edge has proven itself over a large sample. That ceiling stops one suspiciously large Kelly output from overruling your entire bankroll plan. When the formula screams and your history whispers, trust the history.

A step-by-step way to apply Kelly this weekend

Enough theory. Here is how to actually use this without a spreadsheet full of Greek letters.

1. Estimate the true probability honestly. Start from the de-vigged market line as your anchor, then adjust only if you have a specific, nameable reason, and adjust modestly. This is the expected value discipline, and it is the input Kelly lives or dies on.
2. Plug into the formula. f = (bp - q) / b. If it comes out zero or negative, you are done, do not bet. That is Kelly protecting you from a bad number.
3. Cut it to a fraction. Take a quarter or a half of the result. This is the single most important step, because it protects you from your own overconfident estimate.
4. Sanity-check against your unit. If fractional Kelly says something wildly larger than your normal 1% to 2% unit, distrust your probability estimate before you trust the big stake. An enormous Kelly stake almost always means an inflated edge, not a golden opportunity.
5. Adjust for the slate. If you have several live bets, size each smaller, and treat correlated bets as one.

Do this a few times and you will notice something reassuring: on realistic edges, fractional Kelly keeps landing near the small flat units that bankroll management already recommended. The two systems agree, which is exactly why small, disciplined sizing is the safe default whether or not you ever open the formula.

Where a picks service fits

The right way for a service to present sizing is exactly what Kelly implies: a stake scaled to a stated edge, not a "max play" hype label. A tool like ParlayScience markets pick cards with a Kelly-based stake next to the edge, which is at least pointing at correct math instead of "lock of the day." The honest move is to treat any suggested Kelly stake as full Kelly on someone else's edge estimate, then apply your own fraction, because their edge estimate is uncertain just like yours, and you should keep the safety margin. You can see how ParlayScience presents its stakes on Whop and judge it against the review. Never let any stated stake talk you into betting money you cannot lose.

What Kelly gets you that flat betting does not

If quarter Kelly lands near flat small units anyway, why bother with the formula at all? Fair question, and the answer is precise: Kelly earns its keep when your edges genuinely vary and you can estimate them.

Flat betting treats every bet the same, which is safe but leaves growth on the table when some of your bets are much stronger than others. A bettor who truly knows that one play carries a 6% edge and another a 1% edge is leaving money behind by betting them the same size. Kelly captures that difference, telling you to lean into the fat edge and tread lightly on the thin one, and over a long run of well-estimated bets that proportional sizing compounds into meaningfully faster growth. That is the real prize: not knowing whether to bet, but knowing how much more to bet the rare great spot than the ordinary one.

The condition attached to that prize is enormous, though, and it is why most bettors should still default to flat units. Kelly only outperforms flat betting if your edge estimates are actually accurate and reasonably calibrated. If you cannot reliably tell a 6% edge from a 1% edge, and most bettors cannot, then Kelly's proportional sizing is just amplifying your estimation errors, and you are better off flat. So the honest rule is a ladder: flat small units if you cannot estimate edge, quarter Kelly if you can estimate it roughly, and half Kelly only if you have a tracked, closing-line-validated history proving your estimates hold up. Earn the right to size up by proving your edge first.

The takeaway

Kelly answers "how much" with two honest numbers, your edge and the odds, and it is the mathematically optimal growth rule. But full Kelly is a loaded gun that produces 50% drawdowns and assumes an edge estimate you do not actually have. So bet a fraction, half or quarter, capturing most of the growth for a fraction of the pain and building a margin against your own overconfidence. For most bettors, quarter Kelly lands right on the small flat units that keep bankrolls alive. The formula is elegant. The fraction is what saves you.

Bet only what you can afford to lose. If gambling stops being fun, it is time to stop. Help is available (in the US, call 1-800-GAMBLER). 21+, where legal.

FAQ

What is the Kelly criterion? The Kelly criterion is a formula published by John Kelly in 1956 that tells you what fraction of your bankroll to bet to maximize long-run growth, based on your edge and the odds. It bets more when your edge and the payout are larger and tells you to bet nothing on negative-value bets. It is the mathematically optimal growth strategy, with important caveats.

What is the Kelly formula? For a single bet it is f = (bp - q) / b, where f is the fraction of bankroll to bet, b is the decimal odds minus one, p is your estimated win probability, and q is 1 minus p. If the result is negative, the bet is negative expected value and you should not make it.

Why do bettors use fractional Kelly? Because full Kelly produces severe volatility, including 50% drawdowns that are common even when your edge is real, and because it assumes your probability estimate is exactly right when it never is. Half or quarter Kelly captures most of the long-run growth with far less volatility and builds a safety margin against overestimating your edge.

Is full Kelly ever a good idea? Rarely for real bettors. It is optimal only for someone with a perfectly accurate edge estimate and the discipline to bet through massive drawdowns without flinching. Since almost no one has both, betting a fraction is the practical choice, and overbetting is far more dangerous than underbetting.

How does Kelly relate to flat unit betting? Quarter Kelly on realistic edges tends to land close to the 1% to 2% of bankroll that flat unit betting uses, so the two approaches are cousins. If you cannot estimate your edge precisely, betting small flat units is a reasonable, safe approximation of fractional Kelly.

How do I use Kelly when I have several bets on the same day? Scale down. The single-bet formula assumes one bet at a time, so placing full Kelly on many simultaneous bets overbets your bankroll, and correlated bets on the same game make it worse because they can lose together. Cap your total exposure across the slate, treat correlated bets as closer to one position, and reduce your per-bet fraction as the number of live bets rises.