The Kelly Criterion for NBA Betting: Why Fractional Kelly Wins
Loading...
The Kelly Formula Applied to NBA Decimal Odds
I discovered the Kelly criterion during my third season of NBA betting, and I immediately did the worst possible thing with it: I applied it at full strength. Within six weeks, my bankroll had swung between triple its starting value and near zero — twice. The formula works. The problem is that humans do not have the emotional wiring to survive what full Kelly demands.
The Kelly criterion calculates the optimal percentage of your bankroll to wager on a bet with positive expected value. The formula is: Kelly % = (bp – q) / b, where b is the decimal odds minus 1, p is your estimated probability of winning, and q is the probability of losing (1 – p). For UK bettors using decimal odds, b is simply the odds displayed on your bookmaker’s site minus one.
Here is a concrete example. Suppose you estimate a 55% probability that the Boston Celtics will cover a -4.5 spread, and the decimal odds are 1.91. Your values: b = 0.91, p = 0.55, q = 0.45. Plug them in: Kelly % = (0.91 x 0.55 – 0.45) / 0.91 = 0.055 / 0.91 = 6.04%. On a 1,000-pound bankroll, Kelly says to bet 60.40 pounds.
That looks reasonable for a single bet. The problem emerges when you have three or four qualifying bets on the same night, each suggesting 5-7% of bankroll. Suddenly you have 20-25% of your money at risk in a single evening, and a bad night wipes out weeks of progress. Full Kelly maximises long-term growth rate in theory, but the drawdowns along the way are savage enough to make most bettors abandon the system entirely.
Full Kelly vs 1/5 Kelly: Simulation Results Over One NBA Season
A 2020 simulation study by Dotan tested various Kelly fractions on NBA betting data and produced a result that should be required reading for anyone considering this approach. The 1/5 Kelly strategy — betting one-fifth of the full Kelly recommendation — delivered an ROI exceeding 98% over a single NBA season. Full Kelly, by contrast, led to complete bankroll depletion. Not a drawdown. Not a rough patch. Zeroed out.
The gap between those two outcomes is entirely about variance management. Full Kelly assumes your probability estimates are perfectly calibrated — that when you say 55%, you mean exactly 55%, not 53% or 57%. In practice, nobody’s estimates are that precise. A 2% error in your probability estimate can turn a Kelly-optimal 6% bet into an overbetting disaster. Fractional Kelly provides a buffer against estimation error, and that buffer is the difference between survival and ruin.
I ran my own tracking on this across the 2023-24 and 2024-25 seasons. Using 1/5 Kelly, my maximum drawdown was 18% of peak bankroll. Using 1/3 Kelly (a slightly more aggressive fraction), the maximum drawdown was 31%. Using full Kelly on the same set of bets — which I simulated retrospectively, because I am not insane enough to do it live — the drawdown exceeded 70% at one point before recovering. The bets were identical in all three scenarios. Only the sizing differed. That is how powerful stake sizing is relative to selection quality.
The practical lesson: use 1/5 Kelly as your default. If you are confident in your probability estimates (and have the track record to justify that confidence), you might push to 1/4. Going above 1/3 Kelly is gambling with your system’s survival, and no edge is large enough to justify that risk over a full 82-game season plus playoffs.
Implementing Kelly Sizing on UK Betting Platforms
Kelly sizing requires two inputs your bookmaker does not give you: your estimated probability of winning and your current bankroll. That means the system lives in a spreadsheet, not on the platform itself.
I maintain a simple Google Sheet with three columns for each bet: estimated win probability, decimal odds from my bookmaker, and current bankroll. The sheet calculates full Kelly, then divides by five to give me the actual stake. I update the bankroll column after every bet settles — wins increase it, losses decrease it. This means my stakes shrink automatically during losing streaks and grow during winning streaks, which is exactly the behaviour you want from a staking system.
One complication UK bettors face: different bookmakers offer different odds on the same market. If one platform offers 1.91 and another offers 1.95 on the same NBA spread, the Kelly calculation changes. The higher odds increase b, which increases the recommended stake. I always calculate Kelly based on the best available odds, then place the bet at whichever bookmaker is offering that price. This combines Kelly sizing with line shopping, and the two strategies reinforce each other.
About 8% of UK adults bet on sports online, with men comprising the bulk of that figure at 15%. The sheer accessibility of NBA markets across UK platforms means you can implement Kelly sizing across a wide enough sample of bets to let the mathematics work. You need volume for Kelly to deliver its theoretical advantage — a handful of bets per week is not enough. I target 8-12 qualifying bets per week during the NBA season, which gives me roughly 300-400 bets across a full campaign. At that volume, the fractional Kelly edge compounds meaningfully.
A word on responsible implementation. Kelly tells you how much to bet, not whether to bet. If your probability estimate is not supported by genuine research — pace data, injury analysis, matchup assessment — then Kelly will simply help you lose money at a mathematically optimal rate. The formula assumes your edge is real. If it is not, no staking plan will save you. For a broader look at how Kelly fits alongside flat staking and other approaches, I have covered the full comparison in my NBA bankroll management guide.
When Kelly Fails — and What to Do About It
Kelly has two failure modes that every NBA bettor needs to understand before committing to the system.
The first is overconfidence in probability estimates. If you consistently estimate 55% when the true probability is 51%, Kelly will tell you to bet more than you should, and the cumulative overbetting will erode your bankroll over hundreds of bets. The fix is calibration: track your actual win rate by estimated probability band. If you are estimating 55% but hitting 52%, adjust your estimates downward or increase the Kelly divisor.
The second failure mode is correlated bets. Kelly assumes each bet is independent, but NBA bets on the same night’s slate are often correlated. If you bet three NBA spreads and all three involve home favourites, a league-wide trend toward road upsets on that night hits all three bets simultaneously. The drawdown is three times worse than Kelly’s model anticipated. My solution: cap total nightly exposure at 8% of bankroll regardless of what Kelly recommends. If I have five qualifying bets each suggesting 2% of bankroll, that is fine. If I have five suggesting 3% each, I scale them down proportionally to stay under the 8% ceiling.
These guardrails transform Kelly from a theoretically optimal but practically fragile system into something that survives real-world conditions. The mathematics is sound — the challenge is adapting it to the messy reality of imperfect information, correlated outcomes, and human psychology. Fractional Kelly with an exposure cap is not as elegant as the pure formula, but it is the version that actually works over multiple NBA seasons.
Articles
Published by the CourtEdge team.