PropEngine NFL is a from-scratch football model built on public play-level data, not a repackaged odds feed. Here is what drives every projection, how probabilities are checked, and what the board does not do yet.
Each team's pass and rush attempts are projected from its trailing play volume and pass rate, then adjusted by the game's total (more points, more plays) and the spread (trailing teams throw more). Player projections start from that team volume, because opportunity explains most of a player's week-to-week output.
Each player's share of the work — target share, carry share, snap share, air-yards share and WOPR — comes from his season to date. Efficiency rates (catch rate, yards per target, yards per carry, yards per attempt) are blended with position averages because single games are noisy. Target share is trusted quickly; a 2026 backtest of 918 player-weeks showed heavier blending under-projected the top receivers.
Defense-vs-position multipliers adjust each projection for the opponent, pulled toward league average when the sample is small. For outdoor games the kickoff forecast cuts passing production 0.8% per mph of wind above 12 mph (up to 12%) and 3% more in rain. Players ruled Out or on IR are removed, and Questionable players are flagged Q on the board.
Each market uses a distribution that fits its shape: receptions (negative binomial), receiving and rushing yards (gamma), passing yards (normal) and anytime touchdown (Poisson). The probability of going over the line is then calibrated per market against graded results, so a stated 60% should hit about 60% of the time.
Where regulated US books post a prop, the board uses the consensus line and the best available price, and removes the vig across books to get a fair market probability. Expected value is hidden for now: the model is being recalibrated against book lines, and EV returns once enough priced picks have been graded.
When games go final, every prop on the board is graded from official nflverse box scores — wins and losses alike, nothing deleted. The Results page shows hit rate by market and week, calibration (predicted versus actual), the recommended plays, and flat-stake ROI on priced picks.
Same-game legs are not independent. The slip's parlay calculator uses correlations measured from 2023–2026 game logs (for example, a quarterback's passing yards and his top receiver's yards move together at r = 0.44) instead of multiplying leg probabilities.
Projections are estimates and are often wrong. PropEngine is research for adults 21 and over, not betting advice. Gambling problem? Call or text 1-800-MY-RESET.