QB Attempts Sim ?

What it predicts

The full range of pass attempts (and completions) a starting QB is likely to have in one game, not just one number. Each game is played out 20,000 times, snap by snap, and the QB's attempts are counted in every one. Proj is the average, Median the middle outcome, and the 80% range is where his total landed in 8 of 10 simulated games.

How a simulated game works

Both offenses take turns with real down, distance, field position and clock. On every snap the simulator decides run or pass from a model fitted on every NFL play since 2016 (down, distance, field position, score, time left, wind, dome), adjusted for how much this offense likes to throw and how much this defense gets thrown on. A dropback becomes a pass, a sack, a scramble or a penalty. The result of the play (yards, first down, turnover, clock used) is drawn from real plays in the same situation, so third-down conversions, two-minute drills and clock-killing drives happen the way they do on Sundays.

Anchored to the betting market

Each offense's efficiency is tuned until the simulated games average the same score as the Hard Rock spread and total. Game script (who's ahead, by how much, and when) is the biggest swing factor for pass volume, and this keeps it consistent with the market's view of the game. Spread sensitivity in the detail panel shows how many attempts each point of spread or total is worth for this QB.

Early exits

Hard Rock grades the prop if the QB takes a snap, so injuries and blowout benchings count. The simulation includes both. A small in-game injury risk is applied on every dropback (higher on sacks, and higher when he's listed Questionable). In a lopsided fourth quarter a starter can be pulled, at the rate starters actually were in past blowouts. After an exit, attempts go to the backup.

Reading the table

• Line: Hard Rock's attempts line with over/under prices (Hard Rock only), read fresh from the latest capture of this game's line every time the page loads — after kickoff it is the closing line. The projection's P(Over), Fair and Lean were priced at Hard Rock's line when the projection ran; if Hard Rock has moved since, the row shows priced at with that earlier line.
• P(Over): the chance of going over, combining the simulation with Hard Rock's no-vig price. In 2023–2025 backtests the simulation on its own was overconfident against the market (its disagreements were mostly noise), so it is blended at the weight that scored best out of sample (15% simulation for attempts). The simulation's own over/under split is in the row detail.
• Fair line: the half-point line where the simulation is closest to 50/50.
• Lean: how far that blended chance sits from Hard Rock's no-vig price, in percentage points. It is shown as information only. The model has not passed the test below.
• Completions: the same simulation's completions projection against Hard Rock's current completions line (priced at shown when it has moved).

Line-lag watch

When Hard Rock's spread or total moves but its QB attempts line doesn't, the projection is re-priced with that QB's spread/total sensitivity. If the stale line now looks off by enough (at least 0.75 attempts of movement and 4% expected value), it is logged with the price at that moment. After the game the log shows whether Hard Rock's line later moved the same way (closing-line value) and the result. Flags use only odds the site already collects.

Track record and honesty

The model was tested on 2023–2025 the honest way: every season was simulated with a model fitted only on earlier seasons, then compared with Pinnacle's closing attempts line (the sharpest benchmark available) and with this site's earlier props model. Over 1561 QB-games, the model's Brier score was 0.2538 against 0.2490 for the market (lower is better), and a model–market blend scored 0.2489. That did not clear the pre-registered test (beat the market with 95% confidence, be well calibrated, and not trail the older model), so the page runs in shadow mode: projections and probabilities are information, not picks.

When it updates

Data refreshes every morning (overnight play-by-play and the injury report) and the slate is re-scored; a final re-score runs Sunday morning. The model retrains every Tuesday. The line-lag check runs hourly. Use Games to narrow the table to one matchup and click a row for the full distribution.