Stake Sizing with the Kelly Criterion for NHL Betting

Applying Fractional Kelly Stakes to Hockey Bets

Years ago, after reading a chapter on the Kelly criterion in a book that had been recommended to me, I decided I would size every bet using full Kelly. I had a model, the model had positive expectation, and the maths said full Kelly was the growth-maximising stake. Three months later my bankroll was down 47 % despite the model being correct on average. The volatility on full Kelly is brutal, and the gap between “correct on average” and “correct on this particular three-month sample” is enormous in a high-variance sport. I switched to quarter Kelly that autumn and have used it ever since.

This is the article I wish I had read before that summer.

The formula

The Kelly criterion calculates the optimal fraction of your bankroll to stake on a bet given your edge and the price. The formula is straightforward: f = (bp – q) / b, where b is the decimal odds minus one, p is your estimated probability of winning, and q is one minus p.

Worked example. You estimate a moneyline at decimal 2.10 represents a 52 % win probability — three percentage points above the implied 47.6 %. Kelly says f = (1.10 × 0.52 – 0.48) / 1.10 = 0.094 / 1.10 = 8.5 % of bankroll. On a £1,000 bankroll, that is £85 on a single bet.

The maths assumes your probability estimate is correct. If your 52 % estimate is actually 50 %, your edge is zero and Kelly is still telling you to stake 8.5 %. That sensitivity to estimation error is the part that destroys bettors who size off raw Kelly without humility about their own modelling accuracy.

Full versus fractional Kelly

Full Kelly is the mathematically optimal growth strategy if your probability estimates are perfect. They are not. They never will be. The standard professional adjustment is to use a fraction of Kelly — half Kelly, quarter Kelly, even less. Quarter Kelly cuts your expected growth rate roughly in half but cuts your variance dramatically more, which keeps you in the game through the inevitable downswings.

The intuition is that Kelly is a maximally aggressive strategy. It assumes you have infinite tolerance for drawdown, perfect information, and infinite time. None of those conditions hold in real betting. The drawdown tolerance is the key constraint — Kelly users routinely see their bankroll halved or worse during normal variance, and most bettors stop betting (or worse, start chasing) before the upside recovers.

Professional bettors typically nail down 2 to 5 % edge per bet using fractional Kelly and flat percentages of bankroll. Quarter Kelly on a 3 % edge is roughly 0.5 to 1 % of bankroll per bet — far smaller than full Kelly would suggest, far more sustainable in practice.

Applied NHL example

Concrete walkthrough. I model Tampa Bay at home against Columbus and arrive at a 60 % win probability. The market is offering 1.75 decimal — implied probability 57.1 %, vig included. My edge is 2.9 percentage points.

Full Kelly: f = (0.75 × 0.60 – 0.40) / 0.75 = 0.05 / 0.75 = 6.7 % of bankroll. On £1,000, that is £67 on Tampa Bay moneyline.

Quarter Kelly: 6.7 / 4 = 1.7 % of bankroll, or £17 on the same £1,000.

The difference is profound. The full-Kelly stake feels right when you are correct and devastating when you are wrong. The quarter-Kelly stake feels small but sustains across a season of NHL variance. Both strategies have the same expected return per pound staked — quarter Kelly grows the bankroll more slowly because each individual bet stakes less, but it survives long enough to actually realise the growth.

When to skip Kelly entirely

Kelly works when you have a defensible probability estimate. It does not work when your estimate is essentially a guess — which is most of the time for most bettors, in honesty. If your “edge” comes from a gut feeling rather than a model, Kelly is no better than flat staking and arguably worse, because it amplifies the cost of being wrong.

The honest version of Kelly is that it should only be used on bets where your probability estimate is based on something repeatable. A statistical model. A defensible analytical framework. A specific informational edge — a confirmed starting goalie, a fatigue spot, a public-money tell. If you cannot explain why your probability estimate beats the market’s implied probability, Kelly should not size your bet.

For everything else — bets where you have a general feel for the matchup but no rigorous probability estimate — flat staking at 1 to 2 % of bankroll per bet is genuinely the right answer. It is what most professional bettors actually do across most of their portfolio, even when they have Kelly available as a tool.

Drawdown management

The variance of any Kelly strategy is roughly proportional to the Kelly fraction squared. Full Kelly on a 5 % edge has a one-in-three chance of seeing a 50 % bankroll drawdown at some point during a typical betting career. Half Kelly cuts that drawdown risk dramatically — to roughly one in twelve. Quarter Kelly to perhaps one in fifty. The growth rate trades off correspondingly, but the survival probability is what keeps you betting long enough to compound.

The practical drawdown management for hockey bettors goes beyond just the Kelly fraction. Track your portfolio’s correlation. If you are betting five overs on a single night, your effective single-bet exposure is roughly equivalent to one Kelly-stake bet at five times the size. The five tickets are not independent — they correlate through pace of play, refereeing crew composition, and goaltending around the league.

The cleanest fix is to cap your per-day exposure at roughly two to three Kelly stakes total, regardless of how many individual tickets you have placed. The total bankroll exposure on any single calendar day should not exceed what a single full-Kelly bet would have committed — that anchors the variance to something the bankroll can survive.

The discipline that matters

Kelly is not a magic formula. It is a sizing rule that depends on the quality of your probability estimates. Used with humility — fractional, capped, applied only to bets where the estimate is genuinely defensible — it gives you a principled way to size up when the edge is large and size down when the edge is small.

Used carelessly — full Kelly, sized off vibes, applied to every bet without correlation adjustment — it is the fastest way to blow up a bankroll while believing you are doing everything right.

The version I run today: quarter Kelly on model-driven bets, flat 1 % on bets where my edge is qualitative rather than quantitative, hard cap of 3 % bankroll exposure per game and 8 % per calendar day. The bankroll has grown steadily for the last four seasons, the drawdowns have been recoverable, and I have never again ended a summer 47 % down.

For the bankroll-management context that complements this sizing framework, the hockey betting bankroll management guide covers the broader principles of building, protecting and tracking the bankroll Kelly ultimately operates on.

Prepared by the Ice Hockey Betting Tips editorial staff.