Kelly Criterion for NBA Props: Sizing Bets Without Blowing Up Bankrolls
Available in US
Why Stake Size Matters More Than Pick Quality
I once watched a bettor in a north London pub turn a £400 bankroll into £40 over six weeks while picking winners at a clip most professionals would envy. He was right roughly 56% of the time on prop bets where break-even was 52.4%. He had a real edge. He also bet 30% of his bankroll on every wager because he was confident in his picks, and the variance arithmetic destroyed him before his edge could compound.
That story is the entire case for taking bet sizing seriously. The Kelly criterion is the maths that tells you how big a bet should be, given an edge and a price, to maximise the long-run growth rate of your bankroll without taking on ruin risk you cannot afford. It is not a trading system. It is not a way to find better picks. It is the discipline that converts a real edge into actual money over time.
This piece walks through the formula, explains why almost nobody should bet at full Kelly, and shows how fractional Kelly applies to real NBA props.
The Kelly Formula and What Each Term Means
The Kelly formula in its simplest form for a bet at decimal odds is f equals (b times p minus q) divided by b, where f is the fraction of bankroll to bet, b is the decimal odds minus one (the net odds), p is the true win probability, and q is one minus p (the loss probability). It is one line of arithmetic. The challenge is in the inputs.
Take a worked example. A points-prop over priced at decimal 1.91, where your model gives you a true probability of 56% of hitting. The net odds, b, are 1.91 minus 1, which equals 0.91. Your win probability, p, is 0.56. Your loss probability, q, is 0.44. Plug into the formula: (0.91 times 0.56 minus 0.44) divided by 0.91 equals (0.5096 minus 0.44) divided by 0.91, which is 0.0696 divided by 0.91, or about 0.0765. Kelly says bet 7.65% of your bankroll on this prop.
What that 7.65% really means is the stake size that, if every bet you placed had this same edge profile and you reinvested winnings, would maximise the geometric growth rate of your bankroll over an infinite sample. Bet bigger and you grow faster on average but your bankroll volatility goes up faster than your growth rate. Bet smaller and you grow more slowly but your bankroll volatility drops faster than your growth rate.
The maths is elegant in the way information-theory results often are. John Kelly derived it in a 1956 Bell Labs paper that had nothing to do with sports betting — it was about telephone signal noise and information-rate optimisation. The translation to gambling came later, but the underlying logic is identical: a betting decision is an information bet on uncertain outcomes, and Kelly is the optimal trade-off between growth and survival.
Why Full Kelly Is Almost Always Wrong
The maths above gives you the optimal stake under one critical assumption: that your true probability estimate is exactly correct. In real-world prop betting, your estimate is never exactly correct. Your true probability is itself an estimate, made under uncertainty, and that uncertainty cascades through the formula in ways that bettors who run the calculation rarely appreciate.
The Wizard of Odds editorial on this is right when it puts it bluntly: uncertainty is part of the analysis, and pretending you know true probability with precision you do not have is the most expensive mistake in the book. If your model says 56% true probability but the actual true probability is somewhere in a confidence interval from 52% to 60%, the Kelly stake under the lower bound is essentially zero (no real edge) and under the upper bound is much larger. Betting at the point estimate ignores the variance of the estimate itself.
The geometric-growth maths is also unforgiving on the downside. A full-Kelly bet at the optimal stake produces, on average, the maximum long-run growth rate. But the median outcome is below the mean — geometric growth has a fat right tail and a thin left tail. Half of all full-Kelly bettors will, by definition, perform worse than the average across any given sample window. The bettors who fall on the wrong side of variance early can see their bankrolls drawn down to a level where the same percentage stake becomes effectively useless.
The empirical observation behind decades of Kelly literature is that real bettors, even those with verifiable edges, almost never sustain full-Kelly betting. Drawdowns at full Kelly routinely exceed 50% even for genuinely sharp models. Most disciplined bettors find that emotional response to a 50% drawdown disrupts their selection process before the geometric growth can recover. The maths is sound; the application is harder than the maths makes it look.
Fractional Kelly: Half, Quarter, Tenth
The standard adjustment for the problems above is to bet at a fraction of the Kelly stake. Half-Kelly — betting 50% of what the formula recommends — is the most common professional setting. Quarter-Kelly is more conservative, used by bettors with less confidence in their probability estimates. Tenth-Kelly is what serious recreational bettors use when they want the discipline of the framework without the variance of full Kelly.
The mathematical effect of fractional Kelly is well-documented. Half-Kelly captures roughly 75% of full-Kelly’s expected growth rate while halving the variance. Quarter-Kelly captures about 44% of the growth at one-quarter the variance. The trade-off is non-linear in a useful direction — halving your stake size loses you only a quarter of your expected growth, because the variance penalty in geometric growth scales faster than the growth itself.
For most NBA prop bettors I know, quarter-Kelly is the practical setting. The reasoning is straightforward. Probability estimates in NBA props carry meaningful uncertainty — the player’s recent form, the opponent’s recent defence, lineup variance, foul-trouble risk, late-night travel, all introduce noise into the projection. A quarter-Kelly setting builds in a margin of safety that compensates for that noise without sacrificing too much growth.
The bettor in my opening anecdote, applying quarter-Kelly to his 56%-on-1.91 example, would have bet 1.9% of his bankroll instead of 7.65%. On a £400 bankroll, that is £7.60 a bet rather than £30.60. Sustained over six weeks of varying results, the smaller stake would have left him with most of his bankroll intact even through the inevitable bad runs.
Applying Kelly to Real NBA Props
The translation from theory to practice involves three concrete steps. The first is settling on a probability estimate you actually believe. This is the hard part. Most bettors I know who run Kelly seriously have a written process for projecting prop outcomes — model output combined with situational adjustments combined with a reading of the line itself. The probability estimate is the input that determines everything else.
The second step is settling on a Kelly fraction and applying it consistently. Pick half, quarter, or tenth, and stick with it. Switching between fractions based on confidence in any single bet is a recipe for blowing up the framework — Kelly only works as a discipline if the application is mechanical.
The third step is applying the formula across an entire slate of bets, not bet by bet. This is the part most bettors miss. If you are betting six props in a single night, the bankroll exposure across all six is what matters, not any one. Most practitioners cap total nightly exposure at some multiple of the per-bet stake — typical caps are 4x to 8x the single-bet Kelly fraction across a slate. The rationale is correlation: NBA props are not perfectly independent, and a bad league night can sweep multiple bets at once.
The formula also reveals when not to bet. If your probability estimate gives you no edge — if (b times p minus q) is negative — Kelly recommends a negative stake, which in betting terms means do not bet. This is one of the framework’s quietest virtues: it tells you when to walk away with the same arithmetic that tells you when to bet, and the answer comes from the maths rather than from willpower. A broader maths-and-tools framework for NBA props sits naturally alongside Kelly because the no-vig calculation feeds directly into the probability estimate that Kelly uses.
What Kelly fraction do most disciplined recreational bettors use?
Quarter-Kelly is the most common practical setting for serious recreational NBA prop bettors. It captures roughly 44% of full-Kelly’s expected growth rate at about one-quarter the variance, which is the right trade-off for most bettors whose probability estimates carry meaningful uncertainty. Half-Kelly is more aggressive and used mostly by professionals with model validation behind their estimates. Tenth-Kelly is the most conservative end of the spectrum.
Can Kelly recommend negative stakes?
Yes, and that is one of the framework’s most useful features. If your true probability estimate gives you no edge over the implied probability, the Kelly formula returns a negative number — which translates to do not bet. The same maths that tells you when to bet also tells you when to walk away. Treat negative Kelly outputs as a hard pass on that bet rather than a signal to take the other side.
Bankroll Survives Variance, Picks Don’t
The Kelly criterion is the most useful framework in prop betting because it forces a bettor to be honest about three things: the size of the edge, the uncertainty around it, and the cost of being wrong. Most bettors are not honest about any of those, which is why most bettors lose. The bettor who applies fractional Kelly with discipline does not need every pick to be right — the maths handles the variance for him. The bettor who chases pick quality without sizing discipline can be right more than half the time and still go broke. Bankroll survives variance. Picks alone do not.
Created by the ”nba Props Betting” editorial team.
