No gut feelings, no hot takes. The model is market-first, the grading is automatic, and the whole record — good and bad — stays public forever.
For every game, we collect moneyline prices from up to ten sportsbooks. Those prices contain the sharpest available information about who's likely to win — but each includes the book's built-in margin (the "vig"). We strip the vig and average across books to get a clean market consensus probability for each side.
Each sport has its own module that can adjust the consensus using signals specific to that sport (starting pitchers, rest, scheduling spots, and similar). Adjustments are deliberately small — the market is hard to beat, and pretending otherwise is how pick services go broke. New signals run in "shadow mode" first: logged and scored for months before they're allowed to move a number.
The number you see on the board is the model's probability that its favorite wins. A 58% call should win about 58% of the time — no more, no less. The record page shows exactly how often each confidence band actually hits, so you can check our calibration yourself.
Every game the model evaluates gets scored once the game goes final — bet or no bet, win or lose. Grades are append-only: nothing is edited, nothing is deleted, and the history pages are permanent. Published picks additionally track closing line value (CLV): whether the price we published beat the market's final price, which is the honest test of an edge.
The board — every game, every confidence score, every grade — is free, every day. Members additionally get the qualified picks card: the rare games where the model finds the price itself to be wrong, published each morning with the price and book frozen on the card.
Model output is informational and entertainment content, not betting or financial advice. Past results do not guarantee future results.