Stat Analysis
Advanced analytics on player and team performance

End-of-Half Strategy in the NFL

New England Patriots QB Jimmy Garoppolo
Photo: USA Today Sports Images

Guest column by Cole Jacobson

If it happens to the team you're rooting for, you're pissed. If it happens to the team you're rooting against, you get cocky. But no matter what side you've been on, you've seen it happen: the offense using the final seconds of the first half to take a knee or otherwise kill the clock rather than trying to score.

Recent years have seen NFL coaches improve vastly in terms of avoiding old habits of risk aversion, whether that comes down to being more willing to roll the dice on fourth downs or on two-point conversion attempts. But despite this overarching trend, we still regularly see coaches decline to take risks when the game clock approaches halftime. Often we hear the justification of keeping momentum, and often we hear the ridiculous notion of "well, the offensive team gets the ball to open the second half anyway," as if who receives that kickoff should make any impact on a team's goal to maximize its points during the first half. Observing this pattern leads to a natural question: should coaches be more aggressive with the ball in the final seconds of the first half?

To answer this, I started with all of the play-by-play data available via NFLScrapR for the past 11 completed seasons (2009 to 2019). I removed all timeouts, kickoffs, and PATs, and then isolated all plays to occur in the last two minutes of the first half. PATs are removed because before 2018, NFLScrapR had a data entry flaw in which all PATs following defensive touchdowns were omitted. Kickoffs were removed because a kickoff return touchdown does not at all reflect a coach's offensive decision-making, and therefore isn't relevant to our question.

From there, I grouped all plays by the drive on which they occurred, creating a new data frame called PointsScoredOnDriveTable, which includes 4,347 drives. This table includes core information about every offensive drive to start in the final two minutes of a first half, including points scored, time remaining at the start of the drive, how many timeouts the offensive team had at the start of the drive, starting field position, and more.

Summary of the project: in most instances, coaches are deciding to be too passive in late first-half scenarios when it would benefit the offense to stay on the field. However, when it comes to situations when there is only time for one play before halftime, coaches are actually extremely accurate in deciding when to take a shot at the end zone.

The Basics: Are Teams Wasting Opportunities When They Could be Scoring?

For brevity and common sense, we can limit the bulk of the project to drives that start in an offensive team's own territory. If you can't figure out that attempting to score is the proper move when your starting field position is at the 50-yard line or inward, your pure fundamental cowardice goes beyond anything this paper could address. Most NFL coaches agree with this stance; since 2009, there have only been three kneeldowns taken on non-scoring drives in the final two minutes of the first half at the 50-yard line or closer, and one was accidental.

One more key distinction before getting into the analysis: taking a knee isn't the only way for an offense to demonstrate no real intent of scoring. If the offense gets the ball at its own 25 with 40 seconds left and executes one or two run plays before the clock hits zero, everyone watching the game knows that the offense meant to run out the clock. Because of this, I created a new metric, called "Quit" -- any drive was labeled a "Quit" if it started in the offense's own territory in the final two minutes of the half and did not include any pass attempts, offensive timeouts, or offensive touchdowns. (This last modification was because of a Chris Johnson 76-yard touchdown run in 2010 on the opening play of a drive starting with 1:56 left; we can't necessarily say that the offense's goal was to run out the clock there.) It's not a perfect metric, but within reason, if the offense doesn't pass the ball or call timeout once in a two-minute drill situation, we can say there was no real intent of scoring. The numbers back this up; since 2009, there have only been three rushing touchdowns of 50-plus yards in the final two minutes of the first half. (The other two came on drives that also included passing plays, and therefore wouldn't be labeled as "Quits" anyway.)

With these disclaimers out of the way, we can start by looking at the most basic question of all: how often are coaches "quitting" in pre-halftime situations to begin with? I did make a linear model (which I can show to any true statistics nerds interested), but what's much easier to interpret are the following graphs, which display data from the past 11 years. Keep in mind that the "Starting Field Pos" refers to distance from the opposing end zone -- e.g., 90 on the graph means the offense's own 10-yard line.

Looking at four plots simultaneously can be a bit overwhelming. So, here is the same data, but without considering how many timeouts the offense had:

We can see that "quitting" on a drive is very uncommon in most areas of the field when more than 40 seconds are left on the clock. If we use the approximations that result from not considering timeouts, "Quits" occur more than half the time inside a team's own 20-yard line in the 21- to 40-second range, and inside a team's own 40-yard line with 20 or fewer seconds.

Are coaches leaving points on the board in these situations? To answer that, the simplest way is to look at the net offensive points on each drive. This is exactly what it sounds like, taken straight from the scoreboard; i.e., an offensive touchdown goes down as +7, and a field goal as +3, but allowing a pick-six or punt return touchdown would go down as -7. Below are graphs that display the average net points for all "non-quitting" drives in the last two minutes of the half, starting in the offense's own territory:

 

Using these graphs, it looks like a pretty open-and-shut case; in just about every situation, the offense is more likely to gain points than lose them. So, we're done, right?

Unfortunately, it isn't that simple. A large part of the reason that coaches kill the clock is that a turnover in their own territory could be deadly, allowing the opposing team to get quick points before the half even if the play itself doesn't result in a score. Strictly looking at the scoreboard after a turnover doesn't account for this, which is why we have to make a new metric.

A New Metric: Adjusted Net Points

We can already easily quantify how much better a touchdown is than running out the clock, and how much worse throwing a pick-six is than running out the clock. But if the offense punts or turns the ball over on a drive, and that play does not directly result in a touchdown, how much worse off is the offense than it would have been if it had just knelt out the clock? To answer this, we can turn to one of NFLScrapR's unique traits: expected points added, or EPA.

With the help of EPA, we can get a very good estimate of just how harmful those drives are. For turnovers, we can simply look at the average EPA lost on all relevant drives (i.e., last two minutes, own territory) that resulted in a non-scoring turnover, which turns out to be approximately 2.6 points. The process for punts not returned for touchdowns is slightly more complicated, and I will share methodology details to anyone interested, but it results in an average loss of approximately 0.7 points. As a result, we get a new metric, Adjusted Net Points, which lets us evaluate the impact of any outcome:

 

ADJUSTEDNetPoints =
NetPoints - (2.584 * NoScoreTurnovers) - (0.6634 * NoScorePunts)

 

Now, we can finally get to the question we want to answer: when is letting the offense try to score more beneficial than running into the tunnel with the score intact?

 

We can see that the majority of situations are expected to result in a net positive for the offense, even when we're isolating drives that start on the offense's own side of the field. But is it enough of a majority to suggest that coaches should be more willing to take a gamble?

Looking at the graph without timeouts for the sake of simplicity, it's almost never a good idea to "Quit" with 41 seconds or more left on the clock. When it comes to the 21- to 40-second range, the average threshold of where drives become detrimental to the offense is their own 14-yard line. Based on the graph's trajectory, this 14-yard line approximation would also be fairly accurate for the 1- to 20-second range if it had a big enough sample size to register, but there have only been two drives to start inside the 15 with 20 or fewer seconds left that had an attempted pass play since 2009.

Does this match up with the rate at which coaches are "quitting?" We earlier established that "Quits" occur more than half the time inside a team's own 20-yard line in the 21- to 40-second range, and inside a team's own 40-yard line with 20 or fewer seconds to go. Put two and two together, and we have a clear result: coaches are indeed "quitting" on drives more often than they should.

Last Play of the Half: Do I Take the Shot?

For this isolated situation, we can forget the concept of drives altogether. If I only have time left for one single play before halftime comes, do I take the shot and try to score, or accept the score as it is?

I created a data frame called FinalPlayOfHalf_OwnTerritory_Table, which consists of the 1,775 plays since 2009 that started in the offense's own territory and ended the first half. I kept the same "Quit" metric as before, referring to any play that was either a rush attempt or kneeldown as a "Quit" (there has never been a rushing touchdown of 50-plus yards on the final play of a first half since 2009).

We can see that the own 39-yard line is right around when coaches become more likely to pack it in. Is this the right strategy?

Since we're only looking at situations where there's one play to go, we can throw out the "Adjusted Net Points" metric, because turning the ball over doesn't matter unless that turnover happens to get returned to the house. All that matters is whether the offensive team is more likely to score than get scored on for that one specific play, if they are throwing the ball.

 

The smaller sample size here leads to some unconventional-looking data, which is why I included the full-field version too. Both graphs suggest that approximately the offense's own 38-yard line is where a pass attempt on the final play of the half becomes ill-advised. Thus, when it comes to situations where there's only one play to go, coaches are hitting the nail on the head.

For Fun: Most Egregious Kneels and Quits

To give more concrete examples of how this logic can be beneficial, here are some of the most egregious decisions to kneel or "Quit" since 2009. We have to separate the two concepts, because a drive that featured a pass attempt and then a kneeldown wouldn't be classified as a "Quit," since the decision to "Quit" came during the drive rather than at the start.

The most egregious kneeldown by far was the aforementioned "accident." In 2015, the Redskins had the ball at the Eagles' 6-yard line with six seconds remaining in the first half. Kirk Cousins forgot how football worked, and took a knee when he was supposed to spike the ball to stop the clock. The Redskins lost out on 3.33 expected adjusted net points there.

As for kneeldowns that were actually called by the coaching staff, the worst comes from a surprising name: Bill Belichick. In 2016, the Patriots had possession with 1:00 left and three timeouts at their own 34-yard line against the Texans, but chose to take a couple of kneeldowns to enter halftime up 10-0. They lost out on 1.23 expected adjusted net points. This came during Tom Brady's four-game "Deflategate" suspension, with Jacoby Brissett making his first career NFL start. It's fair to say Belichick might have been less passive with his ideal quarterback under center.

If we strictly look at field position, the worst kneel besides Cousins' came in 2010, in Jason Garrett's first career game as the Cowboys' interim coach. The Cowboys had driven to the Giants' 44-yard line, but after a false start and an incompletion put them back to the 49, Garrett chose to take a knee on the final play of the half. The Cowboys lost out on 0.75 expected adjusted points according to a linear model for half-ending plays specifically.

As for the worst "Quits" rather than worst kneeldowns, two nearly identical drives stood out. One came from Jim Harbaugh's 49ers against the Houston Texans in 2013; the 49ers took possession at their own 29 with 1:32 remaining and three timeouts, but ran the ball three straight times for a combined 5 yards and then punted just before the half. The 49ers had 1.12 expected adjusted net points to start the drive. (San Francisco led 21-0 at the time, perhaps contributing to its conservative mindset). Similarly, in 2016, Belichick's Patriots had 1:24 left with all three timeouts and started at their own 31, but their drive consisted of five runs, two of them wiped out by their own penalties, for a combined loss of 8 yards before punting. This was in Jimmy Garoppolo's first career start, during the "Deflategate" suspension. The Patriots, who led 10-7 at the time, started the drive with 1.11 expected adjusted net points.

Possible Sources of Error/Other Comments on Methodology

Like any football analytics project, these findings shouldn't be blindly obeyed in all contexts. Game situations matter, and analytics are used properly when they're helping teams make informed decisions rather than forcing coaches to disregard all other factors at play. For example, player personnel makes a major impact. If a coach isn't particularly confident in his quarterback -- as we saw with Belichick for Brissett and Garoppolo -- he has more justification to take it easy. Similarly, score of the game plays a role as well; if one team is down 31-3 right before halftime, it has incentive to be more aggressive than usual.

On a similar note, the argument of morale can also come into play. Any of us who has ever played a defensive snap has felt that morale boost when we see opponents take a knee right before the half because they doubt they can score. But does this actually make a quantifiable impact on how the game will go after halftime? It would be an extremely small one, if any, and even though I'm all for coaches sending the right message to their players by playing to win instead of not to lose, this project was not the place to approach that debate.

Some readers may be wondering why EPA didn't consist of a bigger portion of the project. The answer is pretty simple; as informative as EPA is in many contexts, the core flaw of EPA as it pertains to this project is that it's based on an offense's expected production entering a drive. For example, say a team starts a possession at the opponent's 10-yard line, with an EPA of 5.0, but then settles for a field goal. This drive would have an EPA of -2.0 points, because it resulted in two points fewer than originally expected. But obviously, choosing to leave the offense on the field was a smarter decision than taking a knee would have been, since the offense still got three more points than they would have if they had taken a knee.

The above graph actually reaches similar conclusions as our earlier graphs (i.e. that "quitting" is not smart with 41-plus seconds left, and that it becomes beneficial at roughly the own 15-yard line with 40 or fewer seconds left). But it shows that being farther from the own end zone often leads to fewer EPA, which is very misleading, because the offense is still more likely to score as its field position gets better. WPA (win probability added) has the same issue, though to a smaller extent.

One more topic worthy of debate comes in the labelling of certain drives as "Quits." Needless to say, there's no one right way to label these correctly. For example, in 2011, Mark Sanchez lost a fumble on the opening play of a drive from the Jets' own 49-yard line with 0:53 left -- this was labeled as a "Quit," but can we guarantee the Jets wanted to run out the clock here? I decided that my criteria of "no pass attempts, no timeouts, no offensive touchdowns" were as accurate as we were reasonably going to get. The more qualifiers you add, the more unnecessarily complicated the project becomes to the average reader.

Looking at the graphs, an unfortunate concession I had to make was to stratify the time remaining into 20-second groups. This is because of the impossibility of having two continuous predictor variables displayed at once while having the graph make sense visually. There's a big difference between 21 seconds and 40 seconds being left on the clock, and the linear models of the project reflect this, but when it came to the graphs, this stratification was a necessity to make the images at all comprehensible.

Thanks for the read, and I hope to hear any positive or negative feedback. I'd like to give a special thanks to Keegan Abdoo of NFL Next Gen Stats and Bailey Joseph of the Oklahoma City Thunder for giving specific R tips.

Cole Jacobson is an Editorial Researcher at the NFL Media office in Los Angeles. He played varsity sprint football as a defensive lineman at the University of Pennsylvania, where he was a 2019 graduate as a mathematics major and statistics minor. With any questions, comments, or ideas, he can be contacted at jacole@sas.upenn.edu and @ColeJacobson32 on Twitter.

UPDATE: Predictive Graphs -- When Should Teams Try to Score?

While informative, the above graphs have a few key issues. One is that they can be difficult to interpret without much of a statistics background. Another is that I had to make the concession of stratifying, or binning, the times into 20-second groups, which unfairly treated 0:21 and 0:40 as the same. This was because the graphs would have looked even more convoluted if there were two continuous predictor variables displayed at once, rather than one continuous and one categorical. Lastly, the graphs look back at past data, rather than directly helping out with future decisions.

To combat these issues, I used linear regression models based on that past data to create predictive graphs. If you're looking for an all-encompassing graph telling you when you should be calling plays with the intent to score, in the style of the New York Times' famed 4th Down Bot, here it is:

While it should be self-explanatory, the blue area means that the offensive team is likely to benefit from trying to score (using the "Adjusted Net Points" formula), while the red area means it is better off going into halftime with the score as is. Along the same lines, we can use this format to display when coaches have usually gone for it in the past, with the blue area meaning that coaches tried to score at least 50% of the time in those situations, and the red representing under 50%:

We can see that the red triangles are way wider in the bottom set of plots. This means that when it comes to field position, coaches are being too passive; in other words, it's almost always a good idea to try to score when your field position is at your own 30-yard line or better, but coaches neglect to do so in some situations. However, the red triangles are also shorter in the bottom plot. This means that when it comes to time on the clock, coaches are actually being too aggressive. Coaches almost always try to score with more than one minute left on the clock regardless of field position, but this is an ill-advised move if the field position is particularly bad, e.g., inside the own 10-yard line.

To summarize, it is still true that coaches are slightly too passive on average -- though it's difficult to tell to the naked eye, there is indeed more blue space in the top set of plots than the bottom. But now we have a more nuanced view of the thought process; coaches overreact to seeing a lot of time left on the clock, while they don't react enough to the prospect of above-average starting field position.

Comments

32 comments, Last at 11 May 2020, 1:53am

1 There is another interesting…

There is another interesting drive you may have overlooked.

https://www.pro-football-reference.com/boxscores/201411160nor.htm#pbp_75
Cincinnati had a late, all-rush, 2nd quarter drive end in a made FG.

The Giants sort of had one in 2012. There was a pass, but it didn't count due to DPI.
https://www.pro-football-reference.com/boxscores/201210070nyg.htm#pbp_92

How did you treat sacks or defensive penalties? Were QB scrambles treated as runs or passes?

 

Re: Belichick -- he sometimes got weirdly passive, even with Brady.
He pulled the same trick against the Dolphins in the last game of 2019.
https://www.pro-football-reference.com/boxscores/201912290nwe.htm

7 Thanks for the read, and for…

Thanks for the read, and for the very good questions. As it pertains to your specific examples, the 2014 Bengals called timeout just before Nugent's field goal attempt, and the 2012 Giants called timeout just before the play that resulted in DPI. Because all drives that had offensive timeouts were not classified as "Quits", neither of these drives were going to be considered "Quits" anyway, even before we consider the made FG or DPI.

For the more broad questions, sacks were treated as pass attempts, and QB scrambles were treated as runs - both of these are how NFLScrapR labels the data (https://github.com/ryurko/nflscrapR).

As for defensive penalties, those on their own would not change a drive's classification. For example, if the offense ran the ball and then benefitted from a 15-yard facemask penalty on the defense, that alone wouldn't make the drive be considered a "No-Quit" (though you would assume the offense would start trying to score after that, if it wasn't already). But if a defensive penalty came during a pass attempt, the pass attempt would stand as a reason for the drive to be considered "No-Quit." So back to your 2012 Giants example, even if they hadn't called timeout right before, that DPI would've still made the drive be a "No-Quit" because the Giants threw the ball.

Belichick's tendencies were an interesting side observation of the project; obviously not the purpose going into it, but a fun nugget to find along the ride. I think 42-year-old Brady might not have been too different from 2016 Garoppolo or Brissett when it came to Belichick's decision making.

2 Looking at adjusted net…

Looking at adjusted net points (https://www.footballoutsiders.com/files/pictures/Guest-041520-5.png)

It seems clear that if I have 0:45 left, 1 timeout, and am at my own 5, simply by wasting my timeout I gain an expected 4 points.

8 Really good observation…

Really good observation regarding the spike on the far right there. The spike exists in the ensuing graph that doesn't consider timeouts too, though it's not quite as big. It's something that I should've addressed in my methodology section, but I got too concerned with keeping word count under control.

What it boils down to is a small sample size. Since it's so uncommon for teams to try to score with such bad field position, there have been only 7 drives since 2009 that involved the offensive team starting inside its own 10-yard line with no timeouts and not "quitting". 5 of these 7 drives started with more than 1:00 left, and 2 of the 7 drives started with exactly 1:00 left. 2 of the 7 drives resulted in touchdowns, including one that started with exactly 1:00 to go. If you're curious, the TD drives were led by HOU Brian Hoyer in a 2015 game against the Colts (started with 1:16 left), and by DAL Dak Prescott in a 2016 game against the Packers (started with 1:00 left).

From the software's point of view, there were only two drives meeting that criteria (no timeouts, inside own 10-yard line) that started with 0:41 to 1:00 to go, and one of those two ended in a touchdown, meaning it sees a 50% chance of scoring in those situations. This is why the graph created that spike, which goes against the obvious truths we know about field position. So to summarize, it's a great example of the dangers of small sample size. Plus, things would've looked different if I happened to use stratifications of "0:40-0:59" and "1:00 or More" instead of "0:41 to 1:00" and "Over 1:00". 

 

 

3 Re: final play Since we're…

Re: final play

Since we're only looking at situations where there's one play to go, we can throw out the "Adjusted Net Points" metric, because turning the ball over doesn't matter unless that turnover happens to get returned to the house.

The "final play" is not always the final play. I'll provide this hilarious example:
https://www.pro-football-reference.com/boxscores/201611200cle.htm

Pittsburgh ran three plays with no time on the clock. An untimed down from the 3 due to defensive holding. An untimed down from the 1 due to DPI (Bell TD). An untimed 2pt conversion.

How were defensive penalties on Hails Mary treated? This occurs more often in the 4th quarter, but isn't completely unknown in the 2nd.

As a Lions fan, this is of some interest, both giving and receiving.
 

9 Nice find with the obscure…

Nice find with the obscure Steelers example. The entire project didn't consider PATs at all, including two-point conversions, so that one wouldn't have mattered. But as for the more broad idea of "final plays" repeating, the way I approached that was by not considering the time on the clock at all. If you're familiar with R at all, here was my method:

pbp_all_ProjectPlays <- pbp_all_ProjectPlays %>% group_by(game_id, game_half) %>% 
  mutate(LastPlayOfHalf = as.numeric(play_id == max(play_id))) %>% 
  ungroup()

In normal English, what that means is that I labeled plays as "Last Play of Half" simply based on being the final one before halftime, whether that came with 0:00 left or 0:35 left. So running multiple plays with 0:00 left wouldn't change the fact that only the last one would count.

On that note, defensive penalties can technically count as the final play of the half, but it's extremely rare, because it would require either a penalty that the offense declined for some reason, or an "after-the-whistle" penalty that is enforced between snaps. It has happened though; a couple of examples are an unsportsmanlike conduct penalty against DAL Byron Jones in a 2017 game against the Chiefs (which came after a half-ending KC Tyreek Hill TD catch), and an unnecessary roughness penalty on OAK Khalil Mack after DEN Trevor Siemian took a knee to end the first half of a 2017 game. 

Since a DPI would realistically never be declined by the offense, DPIs on Hail Mary passes would never be classified as the final play of the half in this project.

4 Instead of Net Points or…

Instead of Net Points or Adjusted Net Points, why not just look at net scoring in the remainder of the half? Check the scoreboard at the start of the possession, and again at halftime. How much did the point differential change by?

It's fine to look at actual scoring instead of expected points since this analysis is already averaging over a huge number of games.

(You could also do something similar using win probability instead of point differential. See what the team's WP would've been at halftime if they ended the quarter with no more scoring, then see what the team's WP actually was at halftime.)

5 Points

Maybe to not ruin the sample by including missed FG's, or the clock running out?

I took the premise of the post to be focusing on the "process"--in other words--do coaches try to "go for it"?

However, seeing the actual results would also be interesting. Obviously, not all FG attempts are created equal. Also, in a sample this large, there are surely times where a TD was scored where the probability was low, and times when a relatively "makeable" FG was missed.

10 This is a very good question…

This is a very good question, and something I considered for a while. I would point to two reasons for not doing it in this version of the project, but I do plan to visit this idea later on.

One reason is along the lines of what Joseph commented. I liked the concept of "process over results", which makes it easier to evaluate if a coach made the right decision based on the information he had available to him at the time, rather than how things ended up turning out. It's easy for us sitting at home to see the offense throw a pick-six and then retroactively say, "see, they should've just ran the clock out," when in reality, the majority of interceptions in these situations come with too little time on the clock for the intercepting team to do anything with it.

But the second, and more important reason to be honest, was about the coding process. NFLScrapR has built-in columns called "score_differential" and "score_differential_post" on each play, which give the score differential from the offensive team's point of view both before and after each snap. While this information is great, the problem is that we don't know whether a team that takes possession late in the first the half is also going to be the team that finishes halftime with possession. Depending on your familiarity with R, I would've had to mutate a new column describing whether the offensive team on any given play was also the offensive team on the final play of the half, and then use a mess of "ifelse()" clauses from there to get to where we want to.

This is doable, and it's still something I plan on doing, but when I combined the extra complicated code with the fact that I already really liked the "adjusted net points" concept, I figured it would make more sense to do in a revision. 

13 They do, if you combine with…

They do, if you combine with the game identifier as well. NFLScrapR has a column called "drive" bulit in already.

The "drive" column on its own wouldn't get what you want. E.G. if you went with:

pbp_all_ProjectPlays %>% filter(drive == 12)), it would look ugly, combining the 12th drives of thousands of different games.

But if you went with pbp_all_ProjectPlays %>% filter(game_id == 2009091302 & drive == 12)), now you would get every play on the 12th drive of the 2009 Eagles-Panthers game.

20 Yes, that could be done…

Yes, that could be done pretty easily. But going with the same "process over results" idea I mentioned earlier, I liked the idea of coming up with an average "points lost due to a turnover", since that's more based on information available at the time rather than looking back after the fact.

But when I get around to the idea of net point differential before halftime, mentioned earlier in these comments, that will address what the opposition did post-turnover on an individual basis.

30 After doing a lot more code,…

After doing a lot more code, I have put together the net scoring in remainder of half idea. I can send pictures of the code and/or graphs to anyone interested, but to summarize, this lined up extremely well with the "Adjusted Net Points" idea, which was rewarding to see.

When being compared to "Adjusted Net Points", using net scoring in remainder of half was slightly more receptive of having good field position, and slightly more critical of having a lot of time left on the clock.

As an example, when we didn't consider timeouts while using "Adj Net", the graph suggested that it was almost always a good idea to try to score with over 1:00 left outside of the own 2-yard line, and it was a good idea to try to score with 0:21-0:40 left outside the own 14-yard line. In contrast, when looking at net scoring for the remainder of each half, the graph approximated that teams should "quit" anywhere inside their own 5-yard line with over 1:00 left, and inside the 10-yard line with 0:21 to 0:40. 

To summarize, "Adj Net Points" ended up almost mirroring the approach taken by looking at net scoring in the remainder of half. For anyone looking for more specific details, feel free to reach out via email (jacole@sas.upenn.edu) or Twitter (@ColeJacobson32).

 

6 One other nuance to consider…

One other nuance to consider: the relative passivity of the opposing coach may also influence the expected adjusted net points for a given drive in these situations. If the broad tendency of coaches were to become more aggressive in these situations, then the net points lost due to a turnover, etc. would likely increase, because the opposing team is less likely to Quit after gaining possession. So I suspect the actual break-even / equilibrium point is probably slightly less aggressive than what the model would suggest. (But still notably more aggressive than coaches' current behavior!)

 

That said, this was a terrific piece. Well-considered, and well-written.

11 Thanks for the read, and the…

Thanks for the read, and the kind words. This is a "galaxy brain" take, one that I admittedly didn't consider at all. The stance I took was to look at how any one coach should address these theoretical situations while treating all of the other 31 coaches as being the same as they are in real life. But it's true that if the collective "Quit Rate" drops across the league, there would be an accompanying shift in all of our data here. A very small one, since we'd be a full degree of separation removed from the decisions currently in question, and it's not like those decisions are usually egregiously wrong as it is. But it is worth noting, so a great point to bring up.

14 On the Kirk Cousins…

On the Kirk Cousins kneeldown, the clock was already stopped following an Eagles timeout. The only purpose spiking the ball could have served was if they wanted to bleed just a bit of clock (and then stop it again afterwards) so the field goal could be the final play of the half, instead of leaving 1-2 seconds and worrying that their kickoff coverage unit might be inept enough to surrender a touchdown in that position.

15 Good point that the Eagles…

Good point that the Eagles had just called timeout. I based the paragraph on a quote from here (http://www.nfl.com/news/story/0ap3000000609339/article/kirk-cousins-takes-knee-leaves-jay-gruden-confused), where Cousins says:

"There was a lot of confusion with what the playcall was and for lack of a better word, I just had a lapse in my decision-making and instinctually took a knee when I should've thrown the ball away to stop the clock. We were very fortunate it didn't end up hurting us [because Washington won the game]."

But as it turns out, since the clock was already stopped, saying that he was supposed to spike the ball to stop the clock is inaccurate. So while he did forget how football worked, I should've worded that more accurately.

16 An oldie

YouTube helped me remember an all time favorite end of half sequence. jets up 10-7 run with 20 seconds left from Pitt 45. Thomas fumbles. Pitt goes for Hail Mary with future Jet Neil oDonnell and completes with three seconds left to around Jet 15. For some reason coslet feels half ran out and is livid that Pitt was allowed to call timeout. He mimicked flipping off the ref on way off the field. 

19 Very well done, and what…

Very well done, and what many "Madden Players" usually yell at their screen and likely agree with: Why not take a shot? Anyway...based on the egregious quits alone, why not add a variable to the model: Tied, Down 3 or less, up 3 or more etc...that may help us zoom in on: too risky vs too safe?

21 Not a bad thought for sure…

Not a bad thought for sure. My thought process was that generally, games aren't considered to be "over" before halftime, so the score shouldn't make a huge impact on a coach's decision making unless it's out of control.

But to look into this idea more, here are the most costly "quits" by a team that was either losing, tied, or leading by 7 or fewer points (besides NE in 2016):

2011: Mark Sanchez's fumble that I mentioned in the project. The Jets led 3-0 over the Bills, and the drive started with 0:53 to go when he lost the fumble on the first play. It's debatable to call this a "Quit" since it's hard to judge intent based on one play, but the Jets had 1.08 expected adjusted net points at the start.

2010: A very similar situation, involving a first-play fumble from Chris Johnson. The Titans trailed 10-3 to the Steelers, and started at their own 23 with 1:45 left. Also might not be fair to call this a "Quit", but the Titans had 0.95 expected adjusted net points at the start.

2009: The Steelers took over on their own 25-yard line with 1:39 to go, trailing the Ravens 14-7. The Steelers ran the ball four straight times (including one that got a first down), killing the rest of the clock. They had 0.85 expected adjusted net points at the start of the drive. Unlike the above two, this one is an inarguable "Quit", and it ended up really hurting the team. Trailing by 7 entering the half, the Steelers went on to lose, 20-17, in OT. They went on to finish 9-7, missing out on a Wild Card spot due to a tiebreaker. Seems like some extra aggression in regulation could've helped out a bit.

25 Cole - brilliantly written…

Cole - brilliantly written. Wonder how stats/recommendations would change if analysis expanded to include two-minute drives in Q4 of tied games. Alternatively/additionally, in Q2/Q4 of playoff games. Do teams tend to be MORE conservative or less conservative in those situations?

Recalling THAT divisional playoff game in Jan 2013 in Denver (full disclosure: Broncos fan), when Joe Flacco and the Ravens tied it up with a Hail Mary in Q4...and then Peyton Manning subsequently (with 0:41 remaining on the clock) took a knee to take it into overtime. Was absolutely furious with John Fox for that decision (as was the rest of that bar), was one those “they absolutely should have tried to win it” moments.

26 Good thought for sure. If we…

Good thought for sure. If we looked at the Q4 group on its own, the sample size might be too small to get anything significant out of it (because the score has to be tied), but lumping those in with our existing Q2 situations might give some cool insight. If I had to guess, I'd assume there's slightly more Q4 aggression because coaches don't wanna rely on the luck of the overtime coin-flip, so it could be worth looking into later.

And yeah, that's the John Fox era for you. Gotta take the bad with the good there.

27 Q4 group

I don't think you would have to limit it to tie score scenarios--you could probably lump in anything where the team with possession is behind by 3 or less. Then they at least have the choice of being aggressive enough to try to obtain a TD, versus milking the clock down and taking the game-winning or game-tying FG. Obviously time plays a factor--but look at the last 2 minutes of last year's MNF opener between NO & HOU. The Texans ended up being too aggressive--they got the TD they HAD to have to take the lead, but gave Brees just enough time to set up the game-winning FG. 

Also, there might be enough drives that start w/1:00+ where you could see a coach being passive and "settling" for a 40+ yd. FG versus being aggressive enough to either get closer or go for the TD.

28 Good thought, especially…

In reply to by Joseph

Good thought, especially regarding the Saints-Texans case study. Crazy game there. But if we expand from tie scores, then it probably becomes a different project altogether. Obviously a team trailing at all in the final 2 minutes of a game, even if only by 2 or 3 points, is going to be throwing the ball unless it starts out with extremely good field position. So this would force us to change our definitions of what constitutes "quitting" vs. "trying to score". 

Still is a very good concept to look into, especially given how important those situations are in real life. So that could be fodder for a future project. 

31 Definitions for those games

Obviously, it would be "settling" versus "quitting," but semantics aside, my idea would be:

Timeouts=20 "adjusted seconds"--in other words, 3 TO's remaining, starting with 1:00=same as starting with 2:00.

Settling=FG attempt of 25+ yards, starting with 2:00 or more on the clock; 35+ yards, more than 1:30 on the clock; 45+ yards, more than 1:00. (If you started with less than 60 "adjusted seconds," you have done well to get a FG attempt that's less than 55 or 60.)

Settling would also include the use of any running play that is not a QB scramble; if you have time to run a running play to get the ball to the appropriate hashmark or center, you have time to run a play to get your kicker closer. (You would also be possessing a TO--so it's not like you have to use the sideline.) If your team is lucky enough to get the ball close to the goal line with extra time left, so that your FG attempt is guaranteed to be the last play of the game (whether tied or trailing), then I suppose that we could also stipulate that any FG attempt of <25 yards does not include settling.

 

32 Yeah, these are good…

Yeah, these are good thoughts for sure. The "adjusted seconds" formula was actually something I considered for this project too, but I ultimately decided to just spell out the actual number of timeouts remaining in the plots. Figured it made sense to limit the amount of jargon where I could. It would have to be a pretty nuanced formula -- e.g., if you have 0:10 on the clock and all 3 timeouts, the timeouts don't really mean that much since you won't have time to use all three, so that's not exactly equivalent to 1:10 with no timeouts. But it is doable.

More broadly, the "settling" idea is a very good one to look into for a future project. Using length of the field goal attempt also would be tough, since a few incomplete shots at the end zone could result in a reasonably long FG attempt, even though the offense actually wanted to score a TD. But like all projects, the proper parameters would have to be set to work around those issues. So I might choose to look into it in-depth during this offseason.