A model that finds real edges is necessary for long-term profit. It is not sufficient. Bettors with profitable edges go broke all the time, because they bet too much on individual plays and run into a bad sequence that wipes them out before the long run can save them.
This article covers bankroll management for +EV NBA bettors specifically: the Kelly Criterion, why fractional Kelly is almost always the right call, how to handle drawdowns, and the unglamorous discipline that separates bettors who survive from those who don't. If you've worked through the rest of our series — the pillar guide, EV calculation, Bet365 specifics, and projection modelling — this is the piece that ties it all to outcomes you actually keep.
This article assumes you understand expected value and edge percentage. If you don't, start with the EV calculation guide before reading this.
Why bet sizing matters more than bet selection
Counter-intuitive truth: among bettors with similar edges, the difference between long-term winners and losers is mostly stake sizing, not pick selection. A 5% edge bettor who stakes 25% of bankroll per bet will go broke faster than the same bettor flat-staking 1%. The expected value is identical; the variance is wildly different; and variance, not EV, is what kills bankrolls.
This is the central insight behind Kelly Criterion: there is a mathematically optimal stake size, given an edge and a price. Bet smaller than optimal and you grow your bankroll more slowly than you could; bet larger than optimal and you grow more slowly AND increase the risk of ruin. The optimal stake maximises long-run logarithmic growth of the bankroll — not single-bet expected value.
The Kelly Criterion formula
For a binary bet (win or lose) with decimal odds and a probability estimate, the Kelly stake as a fraction of bankroll is:
f* = (p × b − q) / b
Where:
f* = fraction of bankroll to stake
p = your true probability of winning
q = 1 − p
b = decimal odds − 1 (the net amount won per unit staked)
Worked example
Tatum over 26.5 at 1.91, our model says 60.0% true probability. p = 0.60, q = 0.40, b = 0.91.
f* = (0.60 × 0.91 − 0.40) / 0.91
f* = (0.546 − 0.40) / 0.91
f* = 0.146 / 0.91
f* = 0.160, or 16.0% of bankroll
Full Kelly says you should stake 16% of your bankroll on this single bet. If your bankroll is £10,000, that's £1,600. Most readers' reaction here should be: that feels like a lot. It is.
Why full Kelly is too aggressive
Full Kelly has three properties that make it impractical for real-world betting:
It assumes your probability estimate is exact. Your model says 60%, full Kelly says stake 16%. But your model could be wrong — maybe true probability is 55%, or 65%. Full Kelly is extraordinarily sensitive to overestimating your edge. If your estimate is high by 5 percentage points, full Kelly will lead you to over-bet and bleed money even though your edge is still positive.
It produces brutal drawdowns. Even when your edge is exactly right, the variance of full Kelly is enormous. A 50% drawdown from peak bankroll has roughly a 50% probability over a long enough sample at full Kelly. Bettors don't typically last through 50% drawdowns — they tilt, overcorrect, or quit.
It assumes infinite divisibility and no limits. Real-world betting has minimum stakes, maximum stakes, line movement between deciding to bet and clicking, and book limits that may stop you executing a full-Kelly stake on a sharp price.
The standard fix is fractional Kelly: bet a fixed fraction of full Kelly. Quarter Kelly (25% of f*) and Half Kelly (50% of f*) are the two most common choices.
Fractional Kelly: the practical choice
Quarter Kelly preserves about 75% of full Kelly's growth rate while cutting variance roughly in half. Half Kelly preserves about 90% of growth at a more modest variance reduction. Both are vastly safer than full Kelly when your edge estimate is uncertain — and your edge estimate is always uncertain.
Worked example, continued
Same Tatum bet. Full Kelly = 16.0% of bankroll. Fractional alternatives:
Half Kelly: 8.0% (£800 on a £10,000 roll)
Quarter Kelly: 4.0% (£400 on a £10,000 roll)
Eighth Kelly: 2.0% (£200 on a £10,000 roll)
Quarter Kelly is our default recommendation for most readers. It tolerates significant errors in edge estimation, produces survivable drawdowns, and still grows the bankroll meaningfully when your edges are real.
If your edge estimates are systematically biased upward by even a few percentage points — and they often are, especially early in your modelling journey — quarter Kelly behaves roughly like Half Kelly with a properly calibrated model. It's a safety margin against your own optimism.
When NOT to use Kelly
Very small bankrolls
Kelly is a fraction. On a £500 bankroll, 4% (Quarter Kelly on a typical edge) is £20. This is fine. But if you find yourself at £100 bankroll, betting £4 per play, the friction of minimum stakes and rounding errors makes Kelly impractical. At small bankroll levels, flat-staking 2–3% of bankroll per bet is simpler and produces similar long-run outcomes.
Bets you don't trust
If your model surfaces a +30% edge bet but your gut says something is off (suspect injury, late line movement against you, weird circumstance), don't bet Kelly. Either pass the bet entirely, or stake a small flat amount that you're prepared to lose. Kelly is for bets where you're confident your edge estimate is calibrated.
Highly correlated bets
Standard Kelly assumes bets are independent. NBA props frequently aren't. Multiple props on the same player, multiple props in the same game, even multiple props on teammates — these covary. If you bet Kelly stakes on five correlated props, your effective bankroll exposure is much higher than 5× the stake of one bet. Either reduce your stakes when betting multiple correlated plays in the same game, or use a multi-bet Kelly extension that accounts for covariance.
Drawdowns and discipline
Even with Quarter Kelly and a real edge, drawdowns are inevitable. A bettor with a +5% edge and bets sized at Quarter Kelly should expect, with high probability, to experience drawdowns of 20–30% from peak bankroll over a year. Most lose less than this; some lose more. Drawdowns of 40%+ happen but are rare.
How to handle them:
Don't increase stakes to recover
The classic gambler's mistake. After a losing run, the temptation is to stake larger to win back what was lost. Don't. Recompute Kelly stakes off the new (lower) bankroll and continue. Increasing stakes during a drawdown is the fastest way to convert a normal drawdown into ruin.
Don't decrease stakes after wins
The mirror image mistake. After winning, the temptation is to lock in profits by reducing stakes. Don't. Recompute Kelly off the new (higher) bankroll and continue. Compound growth requires you to stake proportionally as the bankroll grows.
Recompute the bankroll regularly
Daily or weekly, depending on your volume. The bankroll figure that drives your stakes should be the current bankroll, not your starting bankroll, not your peak bankroll, not the bankroll you wish you had.
Have a stop-loss rule
If your bankroll drops below 50% of starting (or whatever threshold is meaningful for you), stop. Not because the long run has failed, but because that level of loss likely indicates the model is broken, the edges are smaller than you thought, or you've been unlucky enough that you need to verify rather than push through. Stop, audit your closing line value, audit your hit rate by edge bucket, and only restart if the data still supports the strategy.
Practical Kelly workflow for NBA value bettors
A daily routine for someone betting Statz value bets with Quarter Kelly sizing:
Morning: check current bankroll. This is your stake-sizing baseline for the day's bets.
Identify candidate bets from the NBA value bets page. Apply your edge filter (most quantitative bettors use 5–10% minimum).
For each candidate, compute full Kelly using your true probability and the live decimal odds, then take 25% of that figure.
Place the bet, recording: timestamp, line, price, stake, model edge, your true probability.
After tip-off, record the closing line and price.
After the game, record the outcome (win/loss/push).
Weekly: aggregate. Compute hit rate vs implied probability, profit per unit staked, and closing line value. These are the metrics that tell you whether the strategy is working.
The discipline of recording every bet matters more than people realise. Without records, you can't measure your CLV, you can't tell whether your model is calibrated, and you can't distinguish a normal drawdown from a broken edge. Every successful +EV bettor we know keeps meticulous records. Every failed one, in retrospect, didn't.
Beyond Kelly: portfolio thinking
More sophisticated bettors think about their NBA bets not as individual decisions but as a portfolio. The questions become:
What's my total exposure tonight across all bets?
How correlated are my bets? (Same game, same team, same player?)
What's my exposure across stat types? (Am I overweight points, underweight rebounds?)
What's my exposure across edge sizes? (Am I leaning too heavily on a few high-edge bets?)
Without going into full mean-variance optimisation territory, two simple rules go a long way:
Cap total nightly exposure. Most professionals cap their total bankroll at risk on any single night at 10–20% of bankroll, even if Kelly stakes summed across all positive-edge bets would be higher. This is a brute-force correlation hedge — if all your bets go badly on the same night for the same reason (a model bug, a major news event you missed), you don't lose more than 20% of bankroll.
Cap exposure per game. No more than 5–10% of bankroll across all bets in a single NBA game. The bets in one game are inherently correlated — if the pace runs hot, multiple overs hit; if a star gets ejected, multiple lines move against you. Capping per-game exposure forces diversification across games.
The discipline that ties it together
This entire series, summarised in one sentence: build a model that finds real edges, calculate the EV correctly, benchmark against sharp lines, size your stakes proportionally, and let the long run do its work.
None of these steps is glamorous. None will give you the rush of a parlay hit or a fading-the-public win. But the bettors who survive and grow their bankrolls over years do every step, every day, with the same boring discipline. The bettors who don't, however good their reads on individual games, eventually go broke.
If you've made it through all five articles in this series, you have everything you need conceptually to bet NBA props with a real long-term edge. The remaining work is execution — every day, with discipline, recording everything, sizing properly, ignoring short-term variance.
The Statz value bets page handles the model and the edge calculation for you. Your job is the discipline.