An NFL model for point spreads, over/unders and player props, built to answer one question honestly: which factors actually move outcomes, and does any of it beat the betting market?
Walk forward testing: each season was predicted by a model trained only on earlier seasons. The factors were chosen by testing on those same seasons, so these records are a best case, and Vegas is still more accurate on average. The 2026 season board grades every game as it happens.
NFL predictions for every game: projected scores, spread and over/under leans by week and by team, graded as the season goes.
Betting research: does wind affect totals, rain and over/unders, injuries vs the spread, and more, each tested on games the model had not seen.
Public play by play data from nflverse, about 540,000 plays since 2015, turned into pregame ratings: starting QB efficiency, scoring form, success rate, return units, workload from the previous game, wind and injuries to regular starters. Ridge regression projects the margin and total, and a second version starts from the Vegas number and learns what the market misses. Live sportsbook lines are pulled several times a week to compare against.
The pages rebuild automatically from the latest data. Projections are research, not betting advice, and sports betting is not legal in every state.