Rush Attempts Model

How the Rush Attempts Model Works

What it predicts

Official rushing attempts for one game: designed runs, scrambles and kneel-downs, which is the number sportsbooks grade. The headline number is the average of a full count distribution, so 14.3 means "about 14, give or take". The 80% range is where the actual count should land four times out of five. P(Over) is the model's chance the player clears the current Hard Rock line, read off that same distribution.

Two questions, then one number

How many times will the team run? The pregame spread and total (favorites run late to drain the clock, underdogs throw), the team's run rate over expected, pace, scrambles and kneel-downs, how often opponents run on this defense, box counts from FTN charting, and PFF run-blocking and run-defense grades. What share goes to this player? Recent and season carry share, his share when the team was leading vs trailing (weighted by this game's expected script), goal-line and short-yardage role, snap share, depth chart, receiving role, and teammates who are out or returning. A one-stage model fed the same inputs is tested next to the two-stage one, and the blend that predicts best on past seasons goes live.

How it was tested

Against actual carries, not betting lines. Every 2018–2025 season was predicted by a model trained only on earlier seasons, and every 2026 week by a model trained only on earlier weeks. A run goes live only if it beats the last-4-game average, the season average and team carries × carry share on both average error and squared error. Held-out error for RBs and QBs: 2.69 carries per game vs 2.91 for the last-4 average. Full results, plus how often players went over or under their closing lines, are on the backtest page.

Best Bets

A player qualifies when the model's chance of going over differs from Hard Rock's de-vigged price by at least 10 percentage points. The three qualifiers the model is most confident about are the week's Best Bets (★ in the Pick column, where "+14" means a 14-point gap). Since 2023, disagreements of 10+ points have won about 55% at the close. Small disagreements lose to the juice, so they are not shown as picks. Players are skipped when they are Questionable/Doubtful/Out, have no NFL history, practiced limited with no game status yet, or when a teammate the model expects to play has no line. A missing line is how books signal an injury before the report does, so a disagreement there usually means the model is missing the news. Picks are re-checked hourly and freeze at kickoff, and the live record grades them against official carries. Players who don't play are voids. Most picks are unders.

Close record

"vs Close" is how often the player has gone over or under his closing rush-attempts line since 2023. The closing line is the most common line across sportsbooks 45 minutes before kickoff, from The Odds API's historical snapshots. That is the only history available: Hard Rock lines were not captured before September 2026.

When it updates

Data refreshes every morning (play-by-play, injury report, FTN charting), with a final re-score Sunday before kickoff and a retrain every Tuesday. A projection assumes the player suits up. Players listed Out are tagged. Lines are Hard Rock only and appear once Hard Rock posts them, often late in the week.

Data through Oct 19, 2026 · projections refreshed Oct 22, 7:20 AM CT · model trained Oct 20, 11:40 AM CT · Backtest & closing-line study →

★ Best Bets · Week 7 2 qualify

Full record & pick log →

★ Top 3 live 4-2 +1.2u (+20.1%) · 2 pending | backtest Hard Rock 2025+ 58.0% +7.8% n=50 · all books 2023+ 56.9% +4.3% n=216

All qualifiers live 21-15 +3.0u (+8.4%) · 2 pending | backtest Hard Rock 2025+ 58.6% +9.5% n=350 · all books 2023+ 55.0% +3.8% n=2060

★ #1

Kendre Miller

RB · NO vs PIT · Oct 25, 8:30 AM CT

Under 12.5 -125 Hard Rock

Model 70% vs market 52% (edge 18 pts) · projects 10.4 carries

★ #2

Rhamondre Stevenson

RB · NE vs CHI · Oct 22, 7:15 PM CT

Under 10.5 -105 Hard Rock

Model 63% vs market 48% (edge 15 pts) · projects 9.6 carries

Rule: every player whose model chance differs from Hard Rock's no-vig price by 10+ points qualifies; the 3 most confident are the ★ Top 3. Records are 1 unit flat at the Hard Rock price logged at kickoff. Skipped: Questionable/Doubtful/Out, no NFL history, limited in practice with no status yet, or a teammate the model expects to play has no line (books pulling a starter usually means news). Picks freeze at kickoff and are re-checked hourly (last Oct 22, 6:05 PM CT). Most picks are unders, which is where the edge has been. Not a guarantee: see the backtest.

Find a player

Model error2.69 MAE, RB+QB, 8 held-out seasons

Last-4 avg2.91 baseline MAE

80% range 81% historical coverage

Pick a player from the search box or the table to see the projection, the line and the drivers.

Hide projections under 2 Only players with a Hard Rock line Best Bets qualifiers only

Player	Pos	Team	Opp	xCarries ↓	80% range	Team xCar	xShare	HR Line	P(Over)	Pick	vs Close	Top driver

Rush Attempts Model ?