Home › Methodology

Sporlytics Methodology: How the Predictions Are Made

Sporlytics predictions come from two rating systems, one for sports with large scores (NFL, NBA) and one for sports where goals are rare (soccer, NHL, MLB). Both are re-fit from full-season results on every run, use nothing but scores and schedule, and are published next to the no-vig market line so the gap is visible.

Elo ratings: NFL and NBA

Every team starts a season at a rating carried over from last season and regressed toward 1500. After each game the winner takes points from the loser; the amount depends on how surprising the result was and on the margin of victory, using the log-margin multiplier popularised by FiveThirtyEight. Home advantage is a fixed rating bonus per sport. The win probability is the standard Elo formula on the rating gap, the projected margin is the gap divided by a points-per-rating-point scale, and the projected total blends both teams’ recent scoring.

Poisson ratings: soccer, NHL and MLB

Each team has an attack rating (goals scored relative to league average) and a defence rating (goals conceded relative to league average), weighted toward recent games with a half-life of a few months and shrunk toward average for teams with few games. Multiplying home attack by away defence, and vice versa, gives each side’s expected goals. A Poisson distribution on those expected goals produces the probability of every scoreline, which is summed into home win, draw and away win, the most likely score, and the over 2.5 probability. For hockey and baseball, where every game has a winner, the regulation-tie probability is split between the teams in proportion to their strength.

Market probability and edge

The market column is the DraftKings moneyline from ESPN with the bookmaker margin removed: each side’s implied probability divided by the sum of all sides. Edge is the model probability minus the market probability, in percentage points. Gaps under 3 points are shown in grey and should be treated as noise.

What the model does not know

Injuries, suspensions, lineups, travel, weather and motivation. None of it. That is a deliberate limitation: the model is a baseline built from results alone, and the value of publishing it beside the market is that a large disagreement is a prompt to look for the reason, not a bet.

Grading

The last prediction recorded before kickoff is the one that counts. The morning after, each pick is scored as a hit or a miss, and the Brier score (squared error of the probabilities) is recorded for both the model and the market. Games where the model disagreed with the market by 3 points or more are also tracked as flat one-unit bets at the closing odds. Nothing is removed. The full list is on the results page.

Data

Schedules, final scores and lines come from ESPN’s public feeds. Ratings are refit twice a day. Sporlytics is not a sportsbook and takes no bets; it is funded by referral links on the promos page.