Intro
Past posts have explored uses of NFL time average leads as a single-point summary of overall game performance.
Using a simple logistic regression model, I looked at 2021 regular season win luck. Notably, the “luckiest teams” had veteran/high quality QBs.
My intuition: quality QBs perform well in comeback and high leverage situations. The residuals from the single feature model reflect that the direct effect of QBs is significant.
Here, I expand the model to include QB quality, using an approach inspired by Mike Sando’s QB Tiers.
Mike Sando’s QB Tiers
Mike Sando is probably my favorite national NFL writer. He’s one of a few writers that make me a subscriber to the Athletic. His recent tribute to the late John Clayton was tremendous.
Each summer since 2014, Sando has released his QB Tier article, where several NFL team personnel evaluate the league’s QBs using a simple rubric. From the 2021 edition:
TIER 1
A Tier 1 quarterback can carry his team each week. The team wins because of him. He expertly handles pure passing situations. He has no real holes in his game…TIER 2
A Tier 2 quarterback can carry his team sometimes but not as consistently. He can handle pure passing situations in doses and/or possesses other dimensions that are special enough to elevate him above Tier 3. He has a hole or two in his game…TIER 3
A Tier 3 quarterback is a legitimate starter but needs a heavier running game and/or defensive component to win. A low-volume dropback passing offense suits him best…TIER 4
A Tier 4 quarterback could be an unproven player with some upside or a veteran who is ultimately best suited as a backup…— Mike Sando
I really like the tier approach: it aggregates subject matter expert opinion in a straightforward, meaningful way without requiring a false level of precision.
Time Average Lead + QB Quality Model
A summary of the model, fit on regular season games that occured between 2013 and 2020.
- In addition to the game’s time average lead, a team’s Sando QB Tier advantage is included as a feature in a logistic regression.
- I’ve assumed that the following year’s QB Tier is a better indicator of QB quality for a season than their preseason QB Tier.
- This has the tradeoff of reducing the size of the data set.
- If one of the QBs is not evaluated the next season (e.g., retires), that game is not included in the data set.
- I’ve assumed that the following year’s QB Tier is a better indicator of QB quality for a season than their preseason QB Tier.
- For each game, the QB tier rating for QB with the most snap counts played for each team is used.
- For an unranked QB that was the planned opening day starter, I assumed a QB rank of 4.0.
- For an unranked QB that was not a planned opening day starter, I assumed a QB rank of 4.5.
Model Summary
A summary of the coefficients and standard error.
- Time Average Lead:
0.306- Standard Error:
0.018
- Standard Error:
- QB Tier Advantage:
0.337- Standard Error:
0.07
- Standard Error:
End of Season 2021 Tier Rankings
Using (1) the logistic model above and (2) some estimates I made of the Sando QB Tier grade of the QBs that finished the season (i.e., not L. Jackson for BAL, Big Ben rated Tier 3.5 vice 2.64 from 2021 grades), I made a plot of an end of season power ranking based on a team’s season time average lead and QB quality.
The contour lines represent the model’s expected win probability (“exp_wp” in table) against an average team with an average QB. From the plot, think end of season Cleveland Browns led by an injury-plagued Baker Mayfield as roughly representing the baseline.
| rank | team | time_avg_lead | qb_adv | exp_wp |
|---|---|---|---|---|
| 1 | BUF | 5.1 | 1.7 | 89.5 |
| 2 | KC | 4.4 | 2.0 | 88.2 |
| 3 | TB | 4.2 | 1.7 | 86.6 |
| 4 | DAL | 4.8 | 0.7 | 84.7 |
| 5 | IND | 4.6 | -0.3 | 78.9 |
| 6 | GB | 2.0 | 2.0 | 78.3 |
| 7 | LA | 2.8 | 1.0 | 76.6 |
| 8 | ARI | 3.3 | 0.5 | 76.4 |
| 9 | SEA | 1.7 | 1.2 | 71.8 |
| 10 | TEN | 2.4 | 0.5 | 71.0 |
| 11 | NE | 3.1 | -0.8 | 66.6 |
| 12 | LAC | 1.1 | 1.0 | 66.1 |
| 13 | MIN | 1.7 | 0.2 | 64.5 |
| 14 | SF | 1.7 | 0.0 | 62.6 |
| 15 | CIN | -0.1 | 0.7 | 55.4 |
| 16 | CLE | 0.8 | -0.5 | 51.8 |
| 17 | LV | -1.4 | 0.5 | 43.4 |
| 18 | PHI | -0.7 | -0.5 | 40.4 |
| 19 | BAL | -0.4 | -1.0 | 38.6 |
| 20 | NO | -0.4 | -1.3 | 36.6 |
| 21 | MIA | -0.8 | -1.0 | 35.7 |
| 22 | CAR | -0.8 | -1.3 | 33.8 |
| 23 | DEN | -2.0 | -0.8 | 29.5 |
| 24 | CHI | -2.2 | -0.8 | 28.3 |
| 25 | ATL | -3.5 | 0.2 | 27.1 |
| 26 | PIT | -3.1 | -0.5 | 24.6 |
| 27 | WAS | -2.9 | -0.8 | 24.1 |
| 28 | HOU | -2.8 | -1.0 | 23.2 |
| 29 | DET | -4.9 | -0.3 | 17.0 |
| 30 | NYG | -5.0 | -1.5 | 11.5 |
| 31 | NYJ | -6.2 | -1.0 | 9.6 |
| 32 | JAX | -6.6 | -0.8 | 9.3 |
Quick Backtest on the 2021 Season
Model accuracy for 2021 regular season games, using
- 2020 end of season time average lead data
- 2021 preseason Sando QB Tiers.
| Accuracy |
|---|
| 62.1 |
