How the numbers work

The raw material is the NHL's own play-by-play, which logs each shot, hit, block and penalty with a location, and its shift charts, which say who was on the ice second by second. We rebuild from them each morning after the last game goes final.

Odds come from a licensed feed several times a day. We keep each pull and never write over an old one.

Expected goals is the chance a shot goes in, given where it came from and how it happened. The model learned from several seasons of NHL shots and keeps learning as games are played. Add up the chances a team or player gets and you have their xG, which predicts next month's goals better than this month's goals do. Rebounds, rush shots and set plays are labelled from the timing and location of the events before the shot.

Player-card numbers are percentiles against his position, among players with enough ice time to judge. In October, small samples get pulled toward average so a hot week doesn't pass for a career.

Grit & GF rates the stuff box scores skip. Its five parts are physicality, GF (goals and chances from the hard areas), intimidation, chemistry and clutch. We measure each from shift data, pull thin samples toward average, and roll four of them into a 0 to 100 rating against his position. Clutch is still shown on the card but counts for nothing since it failed to repeat in our study.

The methodology requires 15 to 20 games to be accurate. Each part starts at the league average for a player's position and moves toward his own numbers as his minutes add up; around 15 to 20 games, his own play outweighs the starting point. Before that, read a rating as provisional.

Parts with a demonstrated link to scoring also get a goals and wins value. Hits and blocks count in the rating but get zero goals until somebody shows they're worth more. The case for it, and how it could fail, is in Pricing the intangibles. The test came back in Does grit win games?: the rating describes how a player plays and doesn't predict results beyond expected goals, though its chemistry part does, so nothing on the site prices off it.

Games are priced from both teams' scoring and chance rates, adjusted for the starting goalies and home ice, then converted to win, puck line and total probabilities. Changes get tested on seasons the model wasn't fitted on before they go live, and the report card shows how close the probabilities came.

A prop projection starts from his expected ice time and his own rates, adjusts for the opponent, and comes out as a full distribution, so we can price 2.5 shots or 4.5 with the same math. Shots, points, assists, blocks and saves each have their own model and their own test.

We strip the margin out of each book's price to get the market's fair probability. An edge is a gap between that and ours big enough to matter, priced at the best line among books that agree with the consensus. Gaps that look too good to be true, or that come from one book off on its own, stay off the board.

The spotlight flags shots props where something has changed and a line built off his season average may lag: a bigger role, a run of attempts, a soft opponent. A signal only got in after beating the base rate on a season it wasn't built on, and the test reruns weekly.

The short notes on game and player pages are written each morning from the box score and the player's own history. They report what happened and leave the why to you.

Prospect odds come from the 2010 to 2017 drafts: each pick's junior, college and European seasons, and what his NHL career turned into. A prospect today gets matched against past players with a similar line at the same age, in a similar league, from a similar draft slot, after adjusting for how much his team scores and how good his linemates are. October samples lean on last season. Out comes the odds of each tier, his likeliest outcome and a scouting report written from the numbers.

Each projection is stamped with the model version and the time it was made, which is always before puck drop. A new version is a new row; old projections are kept as they were.