Available in USContent
The first drawdown that really taught me bankroll management was eighteen days long. I lost twelve units in a stretch where my analysis was, by every reasonable measure, correct — goalies I’d backed posted .920+ save percentages, totals I’d modelled were within half a goal of the correct line, and my closing-line-value tracker said I’d been right about the market more often than not. Twelve units down. Bankroll intact. I didn’t bust, because I was sized correctly. That’s when I understood what staking actually does — it doesn’t make you profitable, it keeps you alive long enough to be profitable.
NHL bankroll management is the discipline of allocating your total betting capital across individual wagers in a way that survives the inevitable variance of the sport. The crucial difference from case-by-case betting — sizing each ticket by feel — is that bankroll management sets the rules before the emotions arrive, and those rules survive the drawdown that case-by-case sizing doesn’t.
The wider context matters too. The 2024 Gambling Survey for Great Britain put the rate of problem gambling among UK adults at 2.7% using the PGSI 8+ threshold, and the structural feature behind that figure is almost always poor staking discipline rather than poor handicapping skill. A punter who sizes correctly can absorb a long losing run as routine variance. A punter who sizes by gut converts the same losing run into bankroll collapse. The maths is identical; the staking is what diverges.
Unit sizing and why it’s the foundation everything else sits on
A unit is a single, defined fraction of your total NHL betting bankroll. Pick the fraction, write it down, stick to it. Most experienced punters land somewhere between 0.5% and 2% of bankroll per unit, with the more aggressive end reserved for highest-conviction bets and the conservative end being the default for routine plays.
The reasoning behind that range comes from variance maths. NHL betting markets have rough win rates of 50–55% on standard bets at –110 juice, and the variance of a binary outcome stream at that win rate is large enough that even a skilled punter routinely sees ten-bet losing runs. A 1% unit means a 20-bet drawdown removes 20% of bankroll in the worst case; a 5% unit means the same drawdown wipes out the bankroll entirely. The 1% unit absorbs the drawdown and lets the punter keep operating. The 5% unit doesn’t.
The other reason to keep units small relative to bankroll is the long-tail of betting markets. Once you start playing player props, totals at heavier juice, and live in-play markets, the variance compounds. A punter sized at 1% of bankroll across a mix of moneylines, totals and props can comfortably ride out the kind of 15-bet losing streak that crops up two or three times a season. A punter sized at 3% gets uncomfortable in those stretches, and discomfort produces the off-system staking that destroys bankrolls.
One small but underrated rule: define the unit at the start of the season, not retrospectively. A 1% unit on a £2,000 bankroll is £20. If the bankroll grows to £3,000 across the first three months, the new unit at 1% becomes £30. If the bankroll falls to £1,500, the new unit drops to £15. Recalibrating quarterly — or after a defined number of bets — preserves the percentage relationship as the bankroll moves. Refusing to recalibrate is how punters end up sizing at 2% during winning runs and 5% during losing runs, exactly backwards from what survives variance.
The NHL regular season runs 1,312 matches across a six-month window, which means a serious punter places anywhere from 300 to 800 bets in a single campaign. That’s enough volume to compound any sizing error materially. Get the unit right and the rest of the strategy is mostly about preserving that discipline through the months when the maths feels wrong.
Flat staking versus percentage staking and which one survives
Flat staking means betting the same nominal amount on every ticket regardless of how the bankroll moves — £20 per bet, every bet, forever. Percentage staking means betting the same percentage of current bankroll on every ticket — 1% of whatever the bankroll currently shows, recalibrating as the balance moves.
The two systems converge on long-term identical returns when the win rate is stable, but they diverge during drawdowns. Flat staking gets more aggressive as the bankroll falls — a £20 stake on a £2,000 bankroll is 1%, but a £20 stake on a £1,500 bankroll is 1.33%, which means the punter is unknowingly increasing risk as the bankroll shrinks. Percentage staking does the opposite: as bankroll falls, the absolute stake size falls with it, which protects the survival floor.
I run percentage staking, recalibrated weekly, with a hard floor that prevents the absolute stake from dropping below a fixed minimum even if the bankroll halves. The floor is there because a percentage system without one keeps shrinking your stake forever, which means you never recover the drawdown even if you start winning again — the absolute stakes have become too small to matter. The hard floor sits at 0.5% of original starting bankroll, which is my “survival mode” stake when the percentage system would otherwise push me lower.
Some experienced punters use modified flat staking with a periodic reset. Bet £20 per ticket until the bankroll has moved by 20% in either direction, then recalibrate to a new £24 (for +20%) or £16 (for –20%) per ticket. This is a cleaner middle path than pure flat or pure percentage, and the discipline of triggering the reset only at fixed thresholds eliminates the day-to-day variability that erodes adherence to the system.
What I won’t touch is progressive staking — Martingale, Fibonacci, or any chase-the-losses sequence. The maths is unambiguous: progressive systems convert a routine losing run into bankroll collapse, and they offer no compensating gain on winning runs to justify the variance. Anyone selling a progressive staking course is either ignorant of the maths or hostile to your bankroll.
The Kelly criterion and what it actually says about NHL props
The Kelly criterion is a stake-sizing formula derived in 1956 by John Kelly Jr that calculates the bet fraction of bankroll that maximises long-term geometric growth. The formula itself is straightforward: f = (bp – q) / b, where b is the net odds received on the bet (decimal odds minus 1), p is the probability of winning, and q is the probability of losing (1 – p). The output is the fraction of bankroll to stake.
Take a concrete example. An NHL puck line at +1.5 priced at 4/6 (decimal 1.67) where my model says the underdog covers 65% of the time. Plug in: b = 0.67, p = 0.65, q = 0.35. Then f = (0.67 × 0.65 – 0.35) / 0.67 = (0.4355 – 0.35) / 0.67 = 0.0855 / 0.67 = 0.128, or 12.8% of bankroll. That’s the Kelly stake on a bet I think is genuinely 65% to cash at +1.5.
Twelve point eight per cent of bankroll is a huge stake. This is where Kelly gets misunderstood. Full Kelly assumes your probability estimate is exactly right; if it’s even slightly off, you over-stake and bleed bankroll faster than the maths predicts. Almost no serious bettor runs full Kelly. The standard adjustment is fractional Kelly — typically quarter-Kelly or half-Kelly — where you stake one-quarter or one-half of the formula output. Quarter-Kelly on the example above is 3.2% of bankroll, which is still on the aggressive side for routine NHL play but defensible for high-conviction bets where my edge is documented.
The Kelly framework’s biggest weakness on NHL props is exactly the probability estimate problem. For moneyline and totals on liquid games, I can construct a reasonable probability estimate from goals-distribution models and historical data. For player props, the probability estimates carry materially larger error bars, and full or even fractional Kelly on a noisy estimate over-stakes systematically. I size player props at flat units rather than running them through the Kelly formula, which keeps the variance contained even if my prop probabilities are off by a few percentage points.
One subtle point about Kelly that gets missed: the formula’s edge depends on knowing the true probability, and the only objective check on whether your probability estimates are well-calibrated is closing line value over time. If your estimates are right, your bets should be at prices that beat the closing line on average. If they’re not, your Kelly inputs are wrong and you’re over-staking even at quarter-Kelly.
Variance, drawdowns, and the long-run reality of NHL betting
Variance is the maths of how widely your results can scatter even when your underlying skill is positive. A punter with a true 55% win rate at –110 juice has an edge of roughly 4.8 percentage points per bet — small but real. The variance around that edge is large enough that a 100-bet sample can produce a win rate anywhere from 44% to 66% just by chance, even with no change in the underlying skill.
That variance translates into drawdowns. A punter expecting 4.8% per-bet edge on a 1% unit will see drawdowns of 15–20 units routinely, 25–30 units occasionally, and 40+ units once or twice across a long career. None of those drawdowns mean the skill has disappeared. They mean variance has done what variance does. The bankroll management discipline is what carries the punter through those stretches with enough capital to keep operating.
The closing line value framework is the only reliable check on whether your results in a drawdown are variance or skill loss. If your bets are still beating the closing line — that is, your prices are still better than where the market closes — your skill is intact and the drawdown is variance. If your closing line value has flipped negative, your skill has actually deteriorated and the drawdown is signalling a real problem with your approach. The full mechanics of tracking and interpreting closing line value are covered in my piece on CLV as the sharp’s truest skill metric, and bankroll management without CLV tracking is essentially flying blind through extended losing runs.
The behavioural side of drawdowns matters as much as the maths. Tilt — the impulsive deviation from your system after a losing run — is the single largest destroyer of bankrolls among otherwise skilled punters. The most effective antidote I’ve found is a hard daily-loss limit: once you’ve lost three units in a single day, you’re done for the day regardless of how many games remain on the slate. That limit isn’t about minimising variance; it’s about preventing tilt from converting a normal losing day into a bankroll-shaking one.
The staking rules I actually live by
Three rules anchor my staking practice and have through nine NHL seasons. First, the unit is 1% of current bankroll, recalibrated weekly, with a hard floor at 0.5% of starting bankroll. Second, single bets max at 2% of current bankroll on highest-conviction plays with documented CLV history; everything else is flat at 1%. Third, daily loss cap at three units, full stop. Those rules don’t make me profitable. They keep me capitalised through the long stretches when the maths feels wrong, and they shut me down before tilt converts a bad day into a bad week. The maths of NHL betting is generous to disciplined punters and brutal to undisciplined ones, and the only thing that separates the two is whether the staking rules survived the drawdown.
What is a sensible NHL betting unit as a percentage of bankroll?
Most experienced punters size their unit between 0.5% and 2% of total bankroll, with 1% being the standard default for routine bets. The conservative end protects the bankroll through long losing runs that are routine in any binary-outcome sport with juice. Sizes above 2% per bet cross into territory where a single 20-bet drawdown can wipe a significant portion of bankroll regardless of underlying skill.
Does the Kelly criterion actually work for NHL props and totals?
Full Kelly works mathematically when your probability estimates are exactly right, but estimation error on NHL props is significant enough that full Kelly typically over-stakes the bet. Most serious bettors use fractional Kelly — quarter or half — to compensate for estimation noise. On player props specifically, where probability estimates carry larger error bars, flat unit sizing is generally more reliable than Kelly until you have documented CLV history confirming the estimates are well-calibrated.
