How this
works.

Each model is fit on regular-season results. It estimates expected goals or points for both sides, then works out the odds of each outcome on the board: winner, margin, over or under.

Those odds get converted to a fair price, the line where a bet would break even if the model were right. The board puts fair next to the market's open and current prices. Where they differ is the whole point.

The NHL model is written up step by step, with its live team ratings and a changelog, at NHL Methods. Every NHL game card links to a game page with our call in plain words and the main reasons behind it. How those numbers have held up is on Track record.

By sport

Edge is how much more likely the model thinks something is than the betting odds say. A 5% edge means the model gives it 5 points more chance than the odds do. Most edges are small, and small ones are often just noise.

A best bet has to pass three checks: a big enough edge, good enough odds, and a bet type I've opened for picks. It gets a grade (A or B) and a size in units. A unit is a fixed, made-up amount used to keep score.

A lean is the side the model prefers when it didn't pass the checks. It's tracked so it can be checked later, but it isn't a bet and doesn't count in the record.

No play means nothing was worth a bet. Not priced means no odds reached the board, so nothing was judged. Once a game starts, its pick stays up for the record but it's no longer live.

The exact bars, and how a best bet differs from a lean, are in Methods.

The first board posted each day is that day's call. Results grades that saved copy against the final scores. Nothing is changed afterwards, so a bad board stays bad in the Archive.

Straight up: did the team the model picked to win actually win. Totals: did the game go over or under the line the way the model said. The picks record counts best bets only, at the odds shown.

The season's profit on best bets stays hidden until a date I picked before the season started. A running total every morning would tempt me to tinker with the model after a bad week. Until then the boards show how many bets have been graded, not how they're doing.

Projected

The model's expected score before the game. Not a prediction of the exact final.

Win %

The model's win probability. The thin bar under each game is the same split in team colours.

Open / Current

The odds when the board was first posted, and the latest odds. The small line under them shows how they moved.

Fair

What the odds would be if the model were exactly right. If the real odds pay more than fair, that's an edge.

Edge

How much more likely the model thinks it is than the odds say, in percentage points.

Best bet

Passed every check. Shown on black with a grade and a size in units.

Lean

The model's side, tracked but not bet. Shown with a dashed border.

Record

Graded only from the boards as they were posted. A "likely range" is where the true result probably sits; if it runs from a loss to a profit, it's too early to tell.

Where the number comes from

On NHL cards, a link to the game page: our call, the main reasons, the goalies and the odds.

Feed stale

The board hasn't updated in over 30 hours, so it may not be tonight's.

B2B

The team played the previous night.

This is a public record of a model. It isn't betting advice and nothing here is a recommendation to bet. The model doesn't place bets and neither do I on its behalf. The units are bookkeeping so the record can be graded.

Nothing here is proven to win. Telling skill from luck takes thousands of graded bets, and the boards show how far along that count is. A good week proves nothing. Neither does a bad one.

If you bet, bet what you can afford to lose, where it's legal, and only if you're of age. If it's stopped being fun, help is free and confidential. US: call or text 1-800-GAMBLER. UK: GamCare, 0808 8020 133. Canada: ConnexOntario, 1-866-531-2600.