How Three picks

Two markets are read from a public odds feed: the moneyline, which is simply who wins, and the over/under on the two teams’ combined score. Both sides of a price add up to more than 100% because the book keeps a margin. We divide that margin out, which leaves the market’s own probability for each side — for a team to win, and for a total to go over or under the posted number.

Today that feed carries one book’s published line per game. When the licensed multi-book feed is switched on, the card will compare many books and publish the best price of the lot. Until then the card names the book the price came from, and claims nothing more.

Separately we work out each side’s chance from the two teams’ season records and home advantage, using the standard log5 calculation. A short record is pulled toward .500 with 8 pseudo-games, so a 1-0 team is not treated as unbeatable.

A record has to exist before it can say anything, so both teams must have played a minimum number of games: 20 in baseball, 15 in basketball and hockey, 4 in football. Until then the games are not candidates at all. That is why football is absent from the card in the opening weeks of its season, and why no amount of it appearing on the board changes that.

An over/under needs a different model, so it gets one. Each team’s scored-per-game and allowed-per-game rates come from the league’s public standings. One side’s projected score is its own scoring rate, raised or lowered by how much the other side allows relative to the league average — the same arithmetic log5 does for winning, done on scoring. The two projections added together are the projected total. A short season is pulled toward the league average by 10 pseudo-games, the same idea as the records model.

In baseball, one man throws most of the game, so when the feed lists a probable starter his season ERA is charged with 55% of the runs his team is expected to allow, and the team’s own rate with the rest. Earned runs are grossed up to all runs, because roughly eight percent of runs allowed are unearned. When no starter is listed, the team rate stands alone. The card prints both starters and their ERAs in the reason.

Turning a projected total into a probability needs the spread of a single game around its own average. We use the published figure per league: about 4.6 runs in baseball, 2.5 goals in hockey, 13.5 points in the NFL, 17 in the NBA. Runs and goals are counts and lean high, so treating that spread as a normal curve is an approximation, not the truth. On a whole number, which can push, both sides are measured past the number, which is exactly what the two posted prices pay.

One correction matters more than any of the above. A model built on season-long averages can sit above or below a whole day’s board at once — in September a baseball model projects more runs than the market posts almost everywhere, because the season average includes the summer. Being high on every game is drift, not an edge. So each league’s projections are shifted by the average distance between the posted lines and our own numbers that day, which leaves only the disagreements that are specific to a game. A league with fewer than 4 priced games that day cannot tell drift from disagreement, so it gets no over/under pick at all. Neither does a league whose season is too young to have scoring rates.

Our number is 65% of the de-vigged market and 35% of the model — the records model on a moneyline, the scoring model on a total. The market is the sharper input, so it carries more weight; the model is what can disagree with it.

A pick makes the card only if our number is at least 55% — a strong chance of being right — and sits far enough above the market’s own number: +2.0% on a moneyline, +3.0% on a total. Confidence and value together. Chalk at a market price never makes it.

The bar is wider on a total on purpose. The records model reads a whole season of won-lost games; the scoring model reads two averages and one pitcher. It knows less, so it has to disagree by more.

Moneylines and totals compete for the same three slots. The three biggest gaps that clear both tests become the card, at most one pick per game, so a card can be three totals, three moneylines or any mix — whichever the arithmetic picked. On a thin day Three posts two, or one. On a day nothing clears, he posts none and says so.

Read the edge column exactly as it is labelled: the distance between our number and the market’s de-vigged number. It is not a profit forecast. The price you actually pay still carries the book’s margin, roughly two points a side, and beating that reliably needs the multi-book best price described in step one.

A posted pick is never edited and never deleted. When the game finishes, the final score grades it Won, Lost or Push, and the closing price is recorded next to the price we published. A moneyline is graded on the winner. A total is graded on the two scores added together, against the number we posted: level with the number is a Push, counted on the record and not quietly dropped. Days with no card are counted too, so the ledger has no silent gaps.