Tom Brady is the most boring quarterback in the league. There’s no joy in watching his handsome, tomato-free grin march up and down the field each season, winning the AFC East in week four and flashing a handful of Super Bowl rings. He’s boring in his destructive consistency, and for spoiled fans of the Patriots, or the Packers, or even the Colts, the tranquil joy of a franchise quarterback can never top the true emotional high of a journeyman Monstaring his way into glory for one fleeting moment.
Consider Ryan Fitzpatrick. No quarterback better exemplifies the joy and sorrow, the shocking highs and devastating lows. In the 2015 season, Fitzpatrick set Jets records for passing touchdowns and turned a 4-12 squad in 2014 into a 10-5 playoff-bound upstart: he then proceeded to throw 3 interceptions and post a season-worst 42.7 passer rating in a playoff-eliminating loss to Rex Ryan and the Buffalo Bills. For this last season, he had three consecutive 400-yard passing games, beat the eventual 13-3 Saints, and then got benched for Jameis Winston.
I’d call him a roller coaster, but that implies regulation and safety. Fitz is a roller coaster designed by a sadistic seventh grader in Roller Coaster Tycoon: you think you’re having a good time until you get launched into the parking lot and take out half of the park guests. He’s a twenty-sided die with only ones and twenties.
Yet, when you look at his career stats, you see a portrait of a perfectly average quarterback. 190 TDs to 148 picks, 81.1 average passer rating, 60% completion percentage. In ten years, a nerdy, younger version of me can check out Pro Football Reference and wonder why, with stats like that, he never stuck around with one team or made more of an impact. How can we inform that intrepid scholar of the beautiful turbulence of a Fitz game?
Enter, the Fitzmagic Score.
The Fitzmagic Score is an attempt at statistically determining the most volatile quarterback in the league. We’re not interested in calling someone the best, or chastising the worst: we’re here to quantify that innate feeling of uneasy hope you get rooting for that career backup who steps up to try and salvage a season and the embarrassing pain of watching the carriage turn back into a pumpkin at midnight.
To do that, I pulled together game logs for over ten years of QB play (up through 9/23/18), filtering out players with less than ten games and with at least 10 pass attempts. We’re avoiding the rest of the 2018 season to try and see if there’s any predictive power with this score in a future post on linear regression. Huge shoutout to Pro Football Reference for their amazing database which powered this analysis.
I pulled together some high level stats to understand the typical play. In this sample, an average passer rating is around 86, with a standard deviation of 27.8. These baselines give a guideline for how to measure volatility, as we’ll ideally show the quarterbacks with higher standard deviations and variance in their play. For example, Tom Brady has an average passer rating of 101.4 and a standard deviation of 24.6: he’s both better than average and has less variance in his performance. Compare that with our boy Fitz, who is at an 81 passer rating and a standard deviation of 27.4 –below-average rating, fairly normal distribution of games.
Here’s a look at the distribution of passer ratings for all quarterbacks in my data set:
A nice, even distribution of values. While we’re here, let’s see how our boy Fitz compared to Brady’s collection of annoyingly perfect gems.
I absolutely love Fitz giving the league the finger with his histogram, while Brady never ever flirts with absolute awfulness. How can we capture this in one metric?
SHOW YOUR WORK
The first pass at this metric only combined passer rating and deviation.. I built a DataFrame of each quarterback and broke out the standard deviation and average passer rating for each quarterback. With that, I calculated the first draft of the Fitzmagic Score by looking at the QB’s deviation over the average deviation, multiplied by the QB’s average passer rating over the overall rating and indexed to 100. This method tries to adjust for below average performers, but at first glance I think I’ve overweighted their poor play. Here are the top ten Fitzmagic scorers with this methodology, and their rough score.
While we’re accurately capturing the quarterbacks with some volatility, we’re giving too much credence to the bad overall players and not really getting what I’m looking for, which is the capacity for greatness and terror. Let’s take a look at incorporating the maximum values (the best game of their career) and subtracting out the minimum values (the valleys of despair) as a new data point. The difference here should better show inconsistencies: if a QB had a perfect rating of 158.3, but then went out and tossed five picks for a 10 rating, the delta of 148.3 is accurately depicting the joy and the sorrow. Conversely, a QB with solid but unspectacular performances would have a low delta, and receive less of a bump in scoring from this methodology.
Try two takes that delta and brings back our standard deviation work to reward volatility in their total career gamelogs. By taking the delta and multiplying by the player’s deviation over the overall deviation, we’re getting somewhere quite exciting. Here are the top players for this calculation:
Much more exciting! We have Geno as our top quarterback (a nation of Jets fans cries out WHY and forgets his perfect passer rating to end the 2014 season and ensure the Jets had a worse draft pick) but see some fun oddities: Nick Foles’ inconsistency is well-represented, and even a legend like Peyton Manning can’t escape the blips of his final season in Denver. How does our namesake fare? Ryan Fitzpatrick has a Fitzmagic score of 136, while Terrific Tom Brady is at 94.
The average Fitzmagic score is 102, and in a positive sign that we’re in the right direction, the two players closest to the average are Alex Smith and Tyrod Taylor, two quarterbacks known for their consistent but unspectacular and conservative playstyle. We also see a nice distribution of values for our scoring system, broken out in the histogram below: