No tags yet.
Should NFL Teams Punt So Often?
Relative to Major League Baseball and the NBA, NFL teams have been slow to adopt advanced analytics for in-game strategies. Ever since reading about the high school coach in Arkansas that never punts (and wins State Championships), I've been curious if such a strategy would apply to the NFL. So, I downloaded a full season of play-by-play data from NFLSavant.com and went to work.
I developed a model that accounts for the probably of making a first down from every distance (yards-to-go), the expected points from each starting point on the field, and the mean net yards of a punt (Mean Punt Yards - Mean Return Yards = 40 yards). First I'll explain how I arrived at each of these model components, and then we'll discuss the results and implications.
Probability of First Down Conversions
One issue I have with the approach taken by the coach from Arkansas is that he assumes an across the board 50% success rate on 4th downs. I wanted to approach the problem more precisely, and take into account the success rate as a function of down and distance. Surely, a team's success rate will vary from 1 yard to go vs. 15 yards to go, so I felt that should be accounted for in the model.
I also did not want to limit the data-set to only 4th downs. I was able to expand my data-set by also including any 3rd down where the team did not go for it on the following 4th down. This expanded my data-set to all "do-or-die" downs, the thought being that strategically speaking, if you plan to punt on 4th down, then 3rd down is "do-or-die" and analogous to 4th down in terms of play calling.
Below are the predicted conversion probabilities as a function of yards to go for a first down. These predictions are based on a logistic regression model, accounting for all do-or-die downs in the 2016 season and the result (first down = success or failure).
Predicted Points from Starting Yard
Ultimately, winners and losers are decided by who has the most points, so the determinant factor of my model is expected points. To calculate this, I coded every drive with a starting yard and result (number of points). I used linear regression to model expected points from this data.
The regression formula plotted below is: Expected Points = 4.779 - (0.037 * YardsFromEndzone).
Final Model
The final model compares expected points for the opponent if the offense punts vs. the difference in the expected points for the opponent if the offense fails on 4th down, and the expected points for the offense if they succeed. This model can be described as:
(Opponents Expected Points IF PUNT) - (Opponents Expected Points IF FAIL - Our Expected Points IF SUCCEED)
If the number on the left is greater than the number on the right, then it is advantageous to go for it, because you are risking conceding more points by punting.
A more detailed description of the model is:
(ScoringProbability from (100-(YardsFromEndZone - 40)) - (((1 - 1stDownProbability) x ScoringProbability from (100 - YardsFromEndZone)) - (ScoringProbability from YardsFromEndZone x 1stDownProbability))
Here is the model results visualized:
Each row of plots represents yards-to-go. The top row is 1 yard to go, the second is 2, etc. Each column represents yards from the endzone. The first column is 99 yards from the endzone, the second is 98, etc. Each plot that falls above 0 expected points (in the green area) represents a situation where you risk giving up more points by punting than going.
This model suggests that from anywhere on the field, it is advantageous to go for it anytime you have 10 or fewer yards to go for a first down. With 11 yards to go, you should go starting from the 38 yard line. With 12 yards to go, the magic yard line is the 33. If you have 13 or more yards to go, the model suggests that you should always punt. From 30 yards in, under most circumstances, kicking a field goal is the right move, so punting is no longer a consideration.
Sample Scenario
To help visualize how this works, I've illustrated the model below, with the most extreme case, 4th and 10 yards to go from your own 1 yard line (99 yards from the endzone).
If you punt the expected points for your opponent would be 3.28. Your expected points would be 0.
If you go for it and fail the expected points for your opponent is 3.46. Your expected points would be 0.
If you go for it and succeed, the expected points for your opponent is 0 and your expected points is 0.41.
So, your net risk for punting is 3.28 points for your opponent, which is higher than your net risk of 3.05 points for your opponent if you go for it. It is important to note here, that this is taking the most conservative possible approach to expected points. This model leaves the expected points from the starting point of the drive (in this case, the 1 yard line). In reality, expected points will increase as the ball moves down the field. However, because this model prescribes such a dramatic shift in strategy, I wanted to keep it as conservative as possible.
Implications
Here is how the current strategy compares to the optimal strategy.
Currently, teams are punting on 4th down 83% of the time. If my model was applied on 4th down and 10 or less, teams would only punt 27% of the time. So, according to my model, NFL teams are not even close to using the optimal strategy on 4th downs.
Final Thoughts
A common question/suggestion I get when discussing this model is "how does the model account for situation and good offenses vs. bad offenses?" My answer is, "it doesn't." The objective of this project was to determine if an overall, league-wide inefficiency existed with 4th down strategy. If a team were to take this on, they would want to calibrate the model to their own offensive probabilities and expected points.
Regarding situation, the model would need to be used consistently over time to pay off. This is true with any model. A common misconception with probabilities is that they either provide for exact predictions or they are useless. The idea with a model like this is that if you use it overtime, you will realize an advantage over time. It is quite possible that a coach fails the first 5 times they go for it on 4th down and lose by 3 touchdowns. That does not mean the strategy is flawed.
Using predictive models takes commitment and the patience to see it through. Ditching the strategy after 5 attempts would not be a large enough sample to know if the strategy works. This is the primary reason why a data-driven approach to sports (and business) is so hard for so many to swallow. Coaches and leaders are judged quickly and harshly on their failures. Because of that, the accepted strategy is typically much more palatable than something new, even if it is objectively less than optimal.