Risky Business
EdjSports examines critical decisions and their impact on GWC (Game-Winning Chance).
Photo: USA Today Sports Images

There are many factors that affect the game state of an NFL game and have a direct influence on the Game-Winning Chance (GWC). Football is different, in this way, from many other sports. Each down represents a unique situation that is defined by score, ball position, yards-to-first, clock, timeouts and the match up of the teams’ strengths and weaknesses. It is like a book of problems where each turn of the page reveals a new game state we have literally never seen before. We may be able to reference a similar situation that has occurred in the past, but it will not properly account for the astronomical number of combinations that define the play at hand.

The GWC of many free-flowing games such as soccer, hockey, and basketball are largely defined by possession, score, clock, and match up. This is similar to football but without the added layer of discrete downs that have the additional variables of ball position and yards-to-first. Consider baseball, which has been a Sabermetrician’s dream since Bill James began publishing his “Baseball Abstract” in the 1970s. The way teams think about line-ups, shifting, pitching rotations, and on-base percentage has revolutionized that game. Despite all of its beautiful complexity, baseball is missing two key ingredients that distinguish it from football: the running clock and variable scoring increments. In baseball, you are generally trying to produce runs and to keep your opponent from producing runs. The utility of those runs may change as a function of the game state, but typically teams will be compelled to improve or protect their score at all phases of the game. Football is different in that points can be scored in groups of 1, 2, 3, or 6, and the impact of that scoring on GWC can change dramatically as a function of the clock and the score. There are times when a field goal is worthless or a touchdown is too much. And as we have seen this season, it can even be correct to reject scoring voluntarily or, conversely, to allow your opponent to score.

As the Pittsburgh Steelers were putting the finishing touches on an impressive second-half comeback against the Colts yesterday, they faced a fourth-and-1 from their own 44-yard line. With a 28-24 lead and only 2:18 separating him from a divisional title, Mike Tomlin decided to do what most NFL coaches would do; he punted the ball. This choice is more complex than it may appear on the surface. When the goal is to improve GWC, every important variable that defines the game state must be considered. Ultimately, the Steelers were able to squelch a game-winning drive by Philip Rivers and the Colts, but according to a detailed simulation by the EdjSports model, they would have won more often, on average, by attempting a first down rather than a punt. At -4.5% GWC, this was one of the largest fourth-down errors of the week and likely went unnoticed as the Steelers were celebrating their victory.

While this error is significant, it is certainly nowhere near the worst we have seen this season. That is reserved for Kliff Kingsbury at -21% GWC. However, this decision by Tomlin serves as an excellent example of the complexity of football analytics and how changing the game state impacts the correct choice.


Impact of Game State on a Critical Fourth Down Decision

  • Steelers lead 28-24 with 2:18 remaining in the game, fourth-and-1 on their own 44-yard line, Colts have no timeouts


GWC Difference of Punt vs. Go When Adjusting for Score Only:


GWC Difference of Punt vs. Go When Adjusting for Yards-to-First Only:


GWC Difference of Punt vs. Go When Adjusting for Game Clock Only:


The simulation is accounting for some fascinating nuance. When leading by 1, it becomes far more correct to attempt the first down than the actual case. This may seem counter-intuitive, but I suspect the model is picking up on the fact that when the Steelers fail on fourth down, the Colts will not necessarily burn all of the clock to get a field goal. After the punt, the Colts will only need a field goal, of course, but now they are more likely to milk the entire clock in doing so. Surprisingly, the model also slightly favors going for it even on fourth-and-3. This illustrates the importance of retaining possession and keeping the ball away from the opposing offense. Finally, it may seem odd that the GWC difference actually drops when adjusting the clock from 1:00 to 0:20, but this is simply a function of the overall GWC approaching 100% for both choices.

It is illuminating to see how small changes in game state variables can affect whole percentage points of GWC. As we often say, just 6% of GWC per game is an expected difference of an entire game over a 16-game season and often the difference between making the playoffs or not.


15 comments, Last at 31 Dec 2020, 1:10am

1 An additional variable?

I suspect Tomlin's thinking may have been influenced by the Steelers' inability to pick up a single yard on multiple occasions throughout the game.  In addition, the defense seemed to be able to thwart the Colts' offense during the second half.  Both of these may be evidence of the 'recency illusion," of course.  But how do you adjust for short-yardage offensive futility or confidence in your defense? 

2 Great comments, and I think…

Great comments, and I think the answer is, "you don't".

Normal human beings will naturally be influenced by the factors you mentioned, the "recency illusion" as you called it.  But stats will say that PIT's true chance of gaining 1-yard, and IND's true chance of scoring a TD once they had the ball back, are better estimated from longer-term data than what had happened in H2 of that game.  

8 Definitely not to the point…

Definitely not to the point of discounting the game. For instance, if Tomlin saw that his OL wasn't getting significant interior push, all of the calculus changes.

Big reason why I don't like models where they "bake in" team strength assumptions. You at least need to fade that over the course of the game.

I'm *very* surprised the decision gets *worse* only leading by 3 or less, where opponents now only need a FG. That's... just weird, because it means the difference has to be a *strong* function of field position: punting has to take over somewhere around 15 yards away from that, where failing means FG range. (note I'm talking about down by 3 or less)

10 "For instance, if Tomlin saw…

"For instance, if Tomlin saw that his OL wasn't getting significant interior push, all of the calculus changes."

I both agree and disagree with this.  On the surface, it makes perfect sense.  That's what coaches are paid to do, recognize what's working and what isn't, and call plays accordingly.

But then another part of me wonders if this isn't just the same as a kicker misses two FGs and the coach deciding he can't trust his kicker this game?

I think the middle ground is if you're missing your best guys and/or are facing a mismatch, you do need to gameplan around that.  But are we sure your OL can't get interior push, or do you just think they can't based on an insignificant amount of data points?  Watch a bunch of goalline stand plays between the same two teams, and you'll see plays where the DL absolutely blows up the play, and plays where the whole DL gets shoved into the end zone, even though the same personnel are on the field both times.

I suspect the long-term data points about the chance of converting a down-and-distance may often be a better predictor of the chance to succeed than what has happened so far in that particular game.


12 "But then another part of me…

"But then another part of me wonders if this isn't just the same as a kicker misses two FGs and the coach deciding he can't trust his kicker this game?"

You walk over to the kicker, ask him "hey, what happened?" and if the answer is "no idea, it felt good, just didn't go there" then, likely, you trust the long term. If the answer is "the wind's just swirling out there, it's not going to be easy" or "leg just doesn't feel quite right" you knock the accuracy back in your head.

We only see a few field goals/XPs but they practice much more. 

"But are we sure your OL can't get interior push, or do you just think they can't based on an insignificant amount of data points?"

It's not really an insignificant amount: you go and ask the guy who's engaging the defenders on every play how he's doing relative to the defenders. The results might only be important on a small number of plays, but the actual contest happens a ton.


3 How does the model adjust…

How does the model adjust for personnel available, coaching, etc? The chances of an average team succeeding against another average team can only tell you so much about the situation in a specific game.

7 Team strengths

The model is tuned to each team's strengths and weaknesses based on weighted pass/rush DVOA splits, adjusted for backup quarterbacks.

5 Adding another point...

Another major point to add to Tomlin thinking: the actual chances of his opponents (Rivers) to score. Colts couldn't produce ANYTHING in the second half. Did he really made a mistake giving them the ball? No, I dont think so. There was much better chance for Rivers to throw a game-ending INT (and he did it twice - one negated by ticky-tack flag and one dropped by Fitzpatrick) then to throw TD. Hence, punt.

p.s. I'll never forget how last year in LJ MVP campaing Tomlin gave him the ball in OT after winning the cointoss. Why? 'Cause LJ coulnd't do anything, and he didn't. So is Colts/Rivers in second half.

11 2nd & 20 at IND 9(1:10 -…

2nd & 20 at IND 9
(1:10 - 2nd) (Shotgun) N.Hines up the middle to IND 11 for 2 yards (M.Hilton).

3rd & 18 at IND 11
(0:26 - 2nd) (No Huddle, Shotgun) N.Hines right guard to IND 15 for 4 yards (S.Tuitt, M.Hilton).

(0:22 - 2nd) Timeout #1 by PIT at 00:22.

So Tomlin doesn't call a time out after the first play, but only does so after the second when there was only 0:22 left.

But hey. He is a great motivator.

15 At that point the sky was…

At that point the sky was falling for the Steelers, and people were honestly considering whether to replace Ben with Mason Rudolph.  I don't know if this entered into Tomlin's thinking, but I think they were better served there by getting to the locker room to regroup.  That said, although I think overall Tomlin is a good coach, he's never been much of a forward thinker.