Using the Kelly Criterion to Size Bets

The Kelly Criterion is a formula that sets each stake as a share of your current bankroll, sized to how big your edge is, and it tells you to bet nothing when there's no edge at all. You'll run into it once flat betting starts to feel crude: every prop calculator has a Kelly box, and every forum thread on bet sizing ends up there. It answers one question, how much to bet. For most prop bettors, my answer is that it shouldn't be making that call yet.

The version sportsbook bettors use is Kelly % = (bp − q) / b. Here b is the net decimal odds (decimal odds minus 1), p is your own estimate of the win probability and q is 1 − p. Circa Sports puts it plainly: if the result is zero or negative, you don't bet. For a simple two-outcome bet this comes to your advantage divided by what the bet pays to one, which is how Wizard of Odds frames it.

It's built to maximize the expected logarithm of your bankroll, which means the fastest long-run growth. The cost is the size of the swings. MetricGate notes that full Kelly can draw a bankroll down 50% or more even when the edge is real. Betting above Kelly is worse than useless: per Wizard of Odds, betting exactly double the Kelly amount gives zero expected long-term growth, and betting more than that gives expected decline.

In calculators. The Odds-API.io Kelly calculator shows a sample $1,000 bankroll with an 8.33% Kelly fraction: $83.33 full, $41.67 half, $20.83 quarter, $10.42 eighth. Seeing all four next to each other is useful, because the full figure is the one that looks reckless.

In a long-odds side. Wizard of Odds works through a bet that wins 20% of the time at 9-to-2. That's a 10% advantage, and Kelly gives 0.1 ÷ 4.5, about 2.22% of bankroll. A big edge on a long price still gets a modest stake, because you lose most of the time.

In the aggressive case. A 65% estimate on an even-money contract gives a full Kelly stake of 30% of bankroll, as Bot for Kalshi shows. Its warning sums up the whole problem: "Overestimating your edge, even slightly, can turn optimal sizing into a path to ruin."

Kelly assumes two things that player props break. The first is that you know p. On a side or total, a sharp market price gives you something to check your number against. On a rebound line or a strikeout prop, your p is a projection, and topendsports notes that estimation error can be bigger than the edge itself. The worked example above shows how that plays out.

The second is independence. Three props from the same game aren't three separate bets, and per Bot for Kalshi, when positions are correlated, "Kelly overestimates the optimal size". The academic work on parlays doesn't help much here. Christopher Long's 2026 paper shows that a parlay leg belongs only if it's worth betting as a single, but it assumes independent legs, so it says nothing about how to size a same-game parlay.

The fairest argument for Kelly anyway is that flat betting wastes your best spots by staking them the same as your thinnest ones. That's true. The answer to it is fractional Kelly, not full Kelly. Wizard of Odds puts the trade at about half the volatility for about a quarter less growth at half Kelly. A peer-reviewed review of betting strategies found that cutting stakes like this is practically necessary rather than optional, with an adaptive fractional Kelly working well across sports. Karl Whelan's UCD course notes add an honest caveat: fractional Kelly isn't mathematically optimal. It's insurance against your own mistakes, and on props you need it.

Kelly is sometimes treated as a system. It isn't. It only answers how much to bet, never which bets to make. Give it a bad projection and it will size the bad bet very precisely.

Stay on flat units until your record shows your probabilities are worth trusting. That means logged bets with closing prices, and closing line value that holds up as the sample grows. A prop tracker with CLV does the logging. None of the sources here gives a minimum number of bets, so don't believe anyone who offers one with confidence. The test is whether your CLV stays positive as the sample grows rather than bouncing around zero.

When you do switch, use quarter Kelly or less, which is the level topendsports recommends when your probabilities are uncertain. Cap your total exposure on one game and cut stakes further for correlated legs. If full Kelly's swings are what tempt you, it's worth reading how to survive a losing streak first, because full Kelly will put you through several.