Late-Game Strategy With the Lead Vincent Verhei 22 Apr 2021, 12:09 pm
Guest column by Cole Jacobson
As Tom Brady entered the huddle with 1:46 remaining in the 2020 NFC Championship Game seeking to reach his NFL-record 10 th Super Bowl, his Buccaneers had two options. The Bucs faced a third-and-4 while comprising on to a 31 -2 6 contribute. Option one: passed the ball and( probably) not get a first down, but force the Packers to burn their final timeout before rendering Aaron Rodgers possession. Option two: put the ball in the air was endeavouring to seal video games, but with the risk of an incomplete pass that would allow Green Bay to get the ball back and hold on to that coveted timeout.
Most of us remember the outcome; Tampa Bay chose to pass and converted on a pass interference penalty on cornerback Kevin King, and the Packers never touched the ball again. Brady aimed up maintaining his seventh Lombardi Trophy two weeks later. But while the Buccaneers’ gamble worked out in this instance, individual anecdotes are not evidence. This raises the question: when one crew has the projectile with the leading late in a football match, is it more effective to run the projectile and kill the clock, or fling the ball to try to prevent the trailing squad from get the ball again? Using data from NFLFastR, I attempted to find out.
I started with all of the play-by-play data available for the past 20 seasons, including playoffs( 2001 to 2020 ). I originally craved a shorter time span to account for passing turn both more common and more efficient in recent years, but decided that maximizing sample sizes was most important. I isolated all scrimmage play-acts to occur in the last two minutes of each game, eliminating kneeldowns. From there, I developed a fixed where the offensive squad was leading by one self-possession( 2,773 play-acts) and a define where the offensive crew was trailing by one owned( 11,090 plays ). I limited data to the last two minutes because the two-minute warning served as a extremely distinct barometer at which play-calling becomes far more influenced by the clock; e.g ., a losing team’s odds of scoring if it gets the ball back won’t be greatly altered by whether it gets possession with 2:50 to go compared to 3:30.
Summary of the project: while it compares most football analytics discourse, running the ball tends to be more beneficial than extending in helping producing teams hold onto that contribute, mainly when the trailing crew is out of timeouts. Extending is even more efficient than running in these situations than it is generally is, due to the defensive tendency to sell out to stop the run when trailing, but this gap in efficiency generally isn’t enough to outweigh the prospect of killing 40 valuable seconds. However, data shows that, if/ when the losing squad gets the ball back, timeouts aren’t very valuable to it. As a outcome, offenses’ primary motivation for running the ball should be to kill clock , not to force the defense to burn its timeouts, because clock subjects far more than timeouts to the losing team.
Is passing even more effective than usual in late-game clock-killing situations?
The first ordering of business is to figure out if passing is more successful in these clock-killing situations than it is at other points in a game. Basic game theory would suggest that this is the case: the protection knows that the offense wants to drain the clock, which signifies the protection will be selling out to stop the run by stuffing the box while playing no-help man coverage or some variation of it, which opens the door for the offense to pass more successfully than it is generally would. And that’s indeed how it plays out 😛 TAGEND
The above maps indicate the rate at which teams producing by one belonging in the final two minutes of a game got a first down or touchdown on any afforded play.( The black brackets represent 95% confidence intervals .) I use NFLFastR’s distinction between “pass” vs. “run, ” which is based on the intent of a play-act rather than its result( i.e ., scrambles and bags are still “pass play-acts, ” even though the projectile wasn’t thrown ). We can see that for any medium- to long-distance to go, extending is drastically more efficient than passing, and even the 1- to 2-yard range is ambiguous based on the small sample size( 25 extends ).
By comparison, here’s how the same table and graph appear if we consider all play-acts throughout an entire play, regardless of clock or whether the offensive crew leads 😛 TAGEND
I scrapped the confidence intervals here, because the sample size is sufficient enough. While it is true here that extending is more effective for any distance besides the 1- to 2-yard assortment, the gaps between passing and operating are much smaller here than in the prior chart. For example, when considering our “Project Plays”( offense producing by one owned in the final two minutes) in the 3- to 6-yard assortment, extending gave a first down/ touchdown on 43.8% of plays, while operating did so on 21.4%. In differ, for all offensive plays in the 3- to 6-yard scope at any point of video games, extending payed a first down/ touchdown on 47.0% of play-acts, while leading did so on 34.2%. The latter gap is still significant, but less so than the former.
If we isolate third downs, which are traditionally the most polarizing in terms of late-game play-call selects, we can verify this further. Note that there have only been 78 passing plays on first/ second downs in our “Project Plays” since 2001. For what it’s worth, those play-acts have been extremely effective: 30 of the 78 resulted in first downs/ touchdowns despite an average of 9.73 gardens to go, and the plays have had an average gain of 7.77 gardens, including penalties.
The above table/ map represent the third downs within our “Project Plays, ” while the following ones represent all third downs at any point of video games 😛 TAGEND
With the exception of the 10 -plus-yard range, which may be impacted by a small sample size( 52 moves in the “Project Plays” group ), we discover a similar dislocation here to our first information and data. This is particularly flagrant in the 3- to 6-yard scope; in the “Project Plays, ” we had an 18.8% possibility of first down/ touchdown with a drain and 43.4% with a pass, whereas at any point of video games, we had a 42.3% probability with a control and 47.0% with a pass.
On its own, this is by no means groundbreaking information. Any casual football fan’s intuition would be that running becomes harder in these late-game situations where the protection is losing and selling out to stop the run. But it’s valuable to see the data confirm this assumption, and it supplies a helpful first step for our overall project.
How much does occasion remaining impact the trailing team’s odds of scoring?
If our concern was only to learn which of extending or extending was more effective, we could wrap it up. But the issue at hand isn’t just about whether passing generates a better possibility at moving the chains than running. Instead, it’s about whether that advantage created by passing the ball is big enough to outweigh the benefits of potentially killing 40 -plus seconds by participating in the ball.
To approach this aspect, we have to figure out just how much those few extra seconds is to assist the losing squad if/ when they get possession back. We’ll introduce the concept of “adjusted time left, ” which involves adjusting the clock for how many timeouts the trailing squad has.( See the bottom methodology section for more on that .)
I developed a new data frame which includes every drive that began in the final two minutes with the offense trailing by one owned. This allows us to dig into how the time remaining( and other variables, such as field position and timeouts remaining) impact the trailing team’s odds of scoring. We use odds of scoring rather than net points per drive( e.g. stimulated field goal countings as +3, pick-six counts as -7) because, in the specific context of a squad trailing by one rating in the final seconds, it doesn’t have any concern for a turnover that allows the defense to score. When a last-second desperation play leads to levels for the defense, it doesn’t harm the offense’s chances of winning any more than a turnover on downs would. In other words, losing by 14 isn’t any different than losing by 7.
Consequently, we’ll dissect the probability of having a scoring drive instead. With this adjustment, we can treat a pick-six the same as we would an incompletion on a Hail Mary to end the game, as shown in this best-fit plot 😛 TAGEND
Field position refers to distance from the opponent’s end zone( i.e ., “7 0” refers to the offense’s own 30 -yard line ). These arcs reflect what we expect: more time on the clock leads to a higher chance of a scoring drive. While the orange pipeline representing drives starting in opponent territory is particularly steep, perhaps due to a smaller sample size of 83 drives, the clock still plays a major role in the other situations too.
Suppose we take a drive starting at the offense’s own 25, represented by the turquoise curve. Having 100 seconds left and no timeouts leads to approximately a 15% opportunity of scoring, whereas having 60 seconds left and no timeouts leads to approximately an 8% probability. A crack of seven% doesn’t sound like much, but when framed in the context that the theoretical drive commencing with 1:40 left has virtually double the scoring likelihood as one starting with 1:00 to go, we recognise the importance.
To go beyond eye-balling the above plot, I made multivariate regression modelings to predict a team’s chances of scoring based on clock, battleground point, and timeouts. Below is a summary of one of those examples, with examples of how it can be applied 😛 TAGEND
This linear model is of the view that a squad starting from its own 35 -yard line, with one timeout left and 1:10 on the clock, would have a 15.5% probability of scoring, compared to an 8.9% possibility if the clock was at 0:30 instead of 1:10. This predictive framework isn’t perfect( needless to say, it’s impossible to have a negative probability of scoring in any context ), but for the most part, the model’s projected wallop of losing out on 40 seconds is extremely similar to what we saw in the best-fit plot, which was directly taken from real results.
However, the model shows that the number of timeouts is not a strong indicator of the offense’s chances of scoring, with a much less significant P-value than the other variables. This is because timeouts aren’t particularly useful for the losing team once it already has the projectile, since plays during a two-minute drill will never use the majority of the play clock. In other terms, the losing team needs its timeouts most when it is on defense, because that’s when it can use them to prevent the leading team from burning 40 -plus seconds. Such a distinction is crucially important: when the losing team has the ball, clock problems drastically more than timeouts do.
Combining it all: for the leading team, what’s ultimately more conducive to winning?
We’ve quantified how much more likely it is for passing to result in a first down/ touchdown than extending, and how much burning game clock play-acts a role in reducing the losing team’s odds of scoring if it gets the ball back. Which of those two peculiarities affairs more for the leading team?
I initially had two approachings to figuring out which of passing or running was more conducive to the leading team maintaining its leading: one based on the actual eventual winners of each game, and one based on NFLFastR’s win probability added( WPA) metric. I passed code for both options, but I will merely share the first one, primarily to save space but also because WPA’s volatility on a play-to-play basis induced the results a bit noisy. While the itinerary of looking at which team won can often be a dangerous one to take in football discourse, since it’s easy to get sucked into the faulty “running more frequently leads to winning” mindset, it’s safer here because we’re isolating situations where the offensive crew have now been gained a contribute in the final two minutes. As a answer, we’re avoiding the “correlation vs. causation” misfortune that can often happen when a squad runs the projectile 20 -plus times in the second half of the year after conducting 27 -7 at halftime.
With that being said, below are a table/ graph displaying the rate at which teams leading by one owned in the final two minutes of a game end up winning that game, based on whether it runs or extends on any third-down play 😛 TAGEND
Though this is antithetical to just about every football analytics project ever stimulated, in all four interval ranges, running the ball has led to a win more often than passing the ball. We must point out that our confidence intervals suggest that there could be some noise, particularly in the 1- to 2-yard range. But still, the fact that moving had given rise to wins more often than extending even after we control for score, down, and clock is noteworthy, because most “running the ball leads to winning” takes don’t account for the fact that extending crews are often trailing.
One way to get more insight here is to stratify by whether the defense has a timeout or not; i.e ., whether the offense is almost guaranteed to be able to kill 40 -plus seconds with a operate play-act. Below is how often teams eventually is to continue to win after third downs in our “Project Plays” group, when the protection had no timeouts at the time of the crack 😛 TAGEND
And here are the same play-acts, except when the defense did have at least one timeout 😛 TAGEND
While we can disregard the 1- to 2-yard range because of its highly small sample size for extends, the other ranges show us an interesting discrepancy. Consider the chart where the defense had zero timeouts: in all four columns, extending has led to wins more often than extending even after controlling for score, down, and clock. The confidence intervals do overlap in every yardage range, but the facts of the case that similar trends( running over passing) exists for each one can’t be ignored. Meanwhile, consider the chart where the defense does have a timeout: moving vs. extending is basically a laundry, with go even being slightly higher in the 3- to 6-yard assortment. This leads us to a conclusion that shapes sense based on our findings about how timeouts don’t vastly help trailing squads when they have the ball: running the ball is beneficial to the leading team when it knows it can kill 40 -plus seconds, but does not appear to have a tangible impact on boost win percentage otherwise.
We’ve established that when the protection has a timeout, there’s far more incentive to air the ball out. Have coaches behaved properly in recognition of this?
Overall, that’s a resounding yes. When defenses have no timeouts remaining on third downs in the final two minutes, pass have been extremely rare. In contrast, when the protection does have a timeout, tutors have been noticeably more willing to air the ball out. Broadly, coaches have lodge to the correct mindset: operated if you know it’ll kill the clock, but don’t be afraid to threw the ball in the air otherwise.
Real-World Application: Back to the 2020 NFC Championship Game
Let’s say you’re sick of all the nerd graphs, and you simply want to look at how any of this can be applied in a real game. We can look back at that third down from the Buccaneers’ win over the Packers. Keep in psyche that the following points amounts are not based on team personnel; i.e ., they treat the 2020 Packers the same as any NFL team from 2001 to 2020.
Tampa Bay faced a third-and-4 from its own 37 with 1:46 left. Suppose that Tampa Bay would have clinched video games with a first down.( Technically, it wasn’t impossible that Green Bay could have still gotten the ball back, but the chances of getting possession at all, let alone scoring a touchdown in that brief time, are so negligible even for a Hail Mary wardaddy such as Aaron Rodgers that we shouldn’t bother .)
In the past 20 years, squads facing a third down with 3 to 5 gardens to go, while contributing by one belonging in the final two minutes, converted on 27.6% of all strives. They passed the projectile on 72.4% of such tries, with a conversion rate of 22.5%, and attempted a pass play( including clambers and bags) on the other 27.6% of plays, with a success rate of 41.2%.
Let’s assume that, if Green Bay had forced a punt, it would’ve gotten the projectile back with 1:30 to go at its own 25 -yard line. They would have a timeout if Tampa Bay hurled incomplete, but would not if Tampa Bay passed the ball.
Teams to get the ball when trailing by one owned, with at least one timeout, from their own 15 – to 35 -yard line and with 1:15 to 1:45 remaining, have scored on that drive 15.9% of the time. Teams in the exact same conditions, but with no timeouts, have scored 12.9% of the time.
If Tampa Bay extends the ball: 41.2% fortune of winning on that play, and an 84.1% possibility of getting a game-clinching stop after a punt, presupposing Green Bay doesn’t burn a timeout. Total win probability: 0.412+( 0.588* 0.841)= 90.7%.
If Tampa Bay moves the ball: 22.5% fortune of winning on that play, and an 87.1% probability of getting a game-clinching stop after a punt, assuming Green Bay burns a timeout. Total win probability: 0.225+( 0.775* 0.871)= 90.0%.
Based on this extremely general calculation that doesn’t account for either team’s strengths or weakness, Tampa Bay’s decision to pass the projectile was the correct one, by a small margin.
Possible Sources of Error/ Other Comments on Methodology
Like any football analytics project, this shouldn’t be blindly obeyed in all possible contexts. Analytics are used properly when they’re help crews construct informed decisions in the moment rather than forcing coaches to disregard all other factors at play-act. For lesson, player personnel has a significant impact. If a tutor especially has respect for the other team’s passing attack, like the Buccaneers facing Rodgers, he has more justification to make sure the fight squad doesn’t get another chance. Similarly, scouting plays a major role too. If one squad has discerned that the opposed defensive coordinator ever mails members of this house when trailing by a rating in the final times, it should exploit that aggressiveness through the air.
Beyond that general disclaimer, one key caveat specific to this project is the unfortunate necessary of going back to 2001. We all understand how much better the NFL collectively is at throwing the projectile than it was 15, or even five, years ago.( Not to mention that older years have more NFLFastR data entryway wrongdoings .) I originally wrote this project to be since 2011 rather than 2001, but ultimately proceeded with the option that would devote more substantial sample sizes. We must keep in our intellects that the “average team” in the eyes of this project’s data is worse at passing than the true “average team” in today’s NFL.
Another important peculiarity to point out is the unfortunate necessity of categorizing every play-act as either a lead or pass. Needless to say , not every play-act call is that black-and-white, especially with the explosion of RPOs in recent seasons. It’s not fair to label every play-act as a pass or run as if the categories are fully binary, but we do the best we can with the information supplied to us.
As for the methodology, here’s an explanation on what “adjusted time left” involves beyond the summary of “clock adjusted to include timeouts.” For situations where the losing crew was on defense, the formula was simple: real clock, plus 40 seconds for each timeout the defensive crew had left. A defense having 1:00 left with one timeout is more or less equivalent to 1:40 left and no timeouts, because the offense will burn all 40 seconds of the play clock when it can. When the losing team was on offense, it was a more complicated formula to approximate how many extra offensive plays a timeout might establish. For instance, having 0:08 left with one timeout is comparable to having 0:15 with no timeouts, in that the offense almost certainly has at least two but no more than three play-acts left. As such, in my formula, a trailing offense having 0:08 left with one timeout leads to 0:15 of “adjusted time” remaining.
Some readers may be wondering why expected levels added( EPA) was not featured. This is because EPA is based on when the next scoring play in a game are likely to be, rather than a team’s expected point differential for the full remainder of a game. In almost any football situation, this distinction doesn’t matter much: a crew that increases its chances of getting the game’s next scoring play-act is almost always also increasing its chances of winning the game. But in this project, it’s actually likely to increase your chances of winning despite decreasing your own chances of scoring again with a clock-killing run for a short gain. EPA is a strong metric when evaluating situations where the offense cares about scoring, which is nearly always the case. But when draining clock is a bigger priority than get levels, the stat is relatively useless.
Cole Jacobson is an Editorial Researcher at the NFL Media office in Los Angeles. He played varsity sprint football as a nose tackle at the University of Pennsylvania, where he was a 2019 alumnu as a mathematics major and statistics minor. With any questions, comments, or theories, he can be contacted via email at jacole @sas. upenn.edu and @ColeJacobson32 on Twitter.
Read more: footballoutsiders.com