How Overline scores a game

Each team carries an offensive and defensive rating built from points (or goals) scored and allowed, plus win percentage. Early-season numbers are shrunk hard toward league average so a 1–0 start does not look like a dynasty. Football and basketball also ingest ESPN’s FPI/BPI power index — a preseason-stable point-margin rating — so week-one picks are not just home-field plus a coin flip.

Expected score is league average plus home attack, minus away defense, plus a sport-specific home-field constant. Soccer, hockey, and baseball draw scorelines from independent Poisson distributions. Football and basketball use a Gaussian margin with sport-calibrated sigma.

When DraftKings (or another book on the ESPN feed) posts a spread and total, that implied score becomes a Bayesian prior. The board you see is a blend — roughly 60–70% market early in a season, drifting toward the ratings as samples pile up.

Every game gets a computer pick: outright winner, and against the spread / total when a number is posted. The same projection feeds a prop board — team totals, both-teams-to-score, combined TDs, and player markets when ESPN lists a passing, rushing, or receiving leader (or a probable pitcher / goalie). Lines sit on the model mean; recent form only tilts the probability, which is how the desk gets a lean without inventing a second book. Closed games are graded against the frozen pre-game pick — not a live-adjusted score — including the previous week’s slate so the winner record is honest.

Live games take the current score and add remaining expected scoring, with variance shrinking as the clock dies. UFC and tennis project win probability, not a score.

Monte Carlo paths (about 2,200 draws, seeded per game so they stay still) build the histograms, cover probabilities, and 80% intervals. The desk brief is optional and user-triggered — it never runs on page load.