Investigating trade-offs made by American football linebackers using
tracking data
Article in Journal of Quantitative Analysis in Sports · July 2023
DOI: 10.1515/jqas-2022-0091
CITATIONS
2
READS
158
2 authors, including:
Eric Alan Eager
Carolina Panthers
33 PUBLICATIONS 285 CITATIONS
SEE PROFILE
All content following this page was uploaded by Eric Alan Eager on 18 March 2024.
The user has requested enhancement of the downloaded file.
CORRECTED PROOF
J. Quant. Anal. Sports 2023; ( ): 1–14
Research Article
Eric Eager* and Tej Seth
Investigating trade-offs made by American
football linebackers using tracking data
https://doi.org/10.1515/jqas-2022-0091
Received October 26, 2022; accepted June 13, 2023;
published online
Abstract: In recent years, the game of football has made a
shift towards being more quantitative. With the advent of
charting and tracking data, player evaluation is able to be
studied from several different angles. In this paper, we build
and refine two novel metrics: Bite Distance Under Expected
(BDUE) and Ground Covered Over Expected (GCOE) for the
evaluation of linebackers in the National Football League
(NFL). Here, we show that these metrics are heavily correlated with each other, which demonstrates the trade-off
linebackers have to make between being aggressive against
the run and being effective when the opposing offense is
using play-action. We also show that these metrics are more
stable than those in the public space. Finally, we show how
these metrics measure deception by opposing offenses.
Keywords: American football; machine learning; tracking
data; trade-offs
1 Introduction
The game of American football (hereafter referred to as football) has undergone a major shift towards being more quantitative. Many of these changes are finding their way into
public-facing media such as newspapers, radio shows and
television sets. Most public-facing uses of football data are
focused on the decision-making processes of teams, either
off of the field (Clark 2019; Rodrigue 2021) or on the field
(Baldwin 2022; Burke 2013), and a pimary example is in the
form of fourth-down decisions. However, through a variety
of new data sources (Football Outsiders 2022;Next Gen Stats
2022; PFF 2022), player evaluation has improved drastically
from past attempts to quantify the game, utilizing public
*Corresponding author: Eric Eager, SumerSports, Palm Beach, USA,
E-mail: eric.eager@sumersports.com
Tej Seth, SumerSports, Palm Beach, USA; and University of Michigan,
Ann Arbor, USA, E-mail: tej.seth@sumersports.com
play-by-play data like expected points added (EPA), charting
data from Pro Football Focus (PFF 2022) and other companies, and most recently x- and y-coordinate tracking data
(Bliss 2022; Eager et al. 2022; Lopez 2020; Next Gen Stats
2022).
In football there are generally two types of plays; run
plays and pass plays. Historically run plays have been the
most heavily-used plays due to the relative safety of trying
to move the ball while carrying it. The famous Ohio State
coach Woody Hayes once said “only three things can happen when you pass (a completion, an incompletion, and
an interception) and two of them are bad” (Life Magazine
1969). Linebackers, and other players playing in the “box”
(players that are lined up near the line of scrimmage and
between the opposing offense’s left and right tackles – or
end players on the line of scrimmage) are players tasked
with stopping both the run and the pass, and are often the
target of deception by opposing offenses. The rise of passing
as the dominant way to move the football has included an
increase in plays that are called play-action passes, which
are passing plays in which the offense fakes a handoff to
a potential ballcarrier before dropping back to pass. These
plays are very successful, and have been so even through
an increase in their usage (Baldwin 2018a; Baldwin 2018b;
Eager et al. 2022; Hermsmeyer 2019). As play-action usage
has increased, our understanding of the linebacker position
has not necessarily grown with it, due in large part to the
trade-offs that are necessary for them to be proficient in
defending both the run and the pass.
In Eager et al. (2022) we introduced three new metrics that combined the charting data from PFF and the x
and y coordinate tracking data from the National Football
League (NFL) in an effort to improve upon player evaluation. One of these metrics, Bite Distance Under Expected
(BDUE), measured how much linebackers and other box
players reacted to play-action passes during the 2017 to 2020
NFL seasons. Measuring how much defenders react to such
a play (whether they “bite” or not) gives us an edge in
evaluating defensive players as well as offensive schemes
and their play callers.
While not biting on play-action passes is ostensibly
a good thing, we conjecture it comes at the expense of
CORRECTED PROOF
2 — E. Eager and T. Seth: Trade-offs by linebackers using tracking data
something else. A player who chooses to stay and not bite
on play-action passes does that at the expense of defending
run plays. This might be rational, as most running plays are
not as valuable as passing plays, but in other cases it might
be an underreaction. Given the power of both charting and
tracking data, we can test this hypothesis.
Thus, in this paper, in addition to making updates to our
BDUE model which we use on play-action plays, we build
a second model for run plays called Ground Covered Over
Expected (GCOE). Specifically, BDUE and GCOE builds on
previous work in the football analytics space using tracking
data, which has mostly focused on the dropback passing
game (Burke 2019; Chu et al. 2020; Deshpande and Evans
2020; Reyers and Swartz 2023). We find that, like BDUE, GCOE
is a stable metric from season-to-season on a player-level
and, as theorized, is negatively correlated with BDUE. For
most players there is a trade-off between being aggressive
stopping the run and being discerning on play-action passes.
We show that players who buck this trend are players who
are able to stop the run effectively and avoid vacating valuable areas of the field on play action. We also show how well
offenses induce or prevent linebacker movement, which
gives us insight into teams, players and play calling schemes
that take advantage of defenses in the NFL.
2 Data
The two main sources of data in this paper are PFF playby-play charting data (PFF) and the NFL’s Next Gen Stats
(NGS) tracking data (Next Gen Stats 2022). We use additional
data, like stadium and weather data, along with NFL Scouting combine (and college pro data). For a reference on the
differences between charting and tracking we recommend
a thorough review by Statsbomb (Burriel and Buldu 2021).
In summary, charting data is generated by humans’ tracking events – features like whether or not a play was a
play-action pass – while tracking data is data generated by
devises like RFID chips in players’ shoulder pads or through
computer vision.
2.1 PFF charting data
PFF’s charting data feeds feature both a blend of subjective
and objective features that can be included in the modeling
process for BDUE and GCOE. We provide justification for
including these features in the model below:
– Outside Zone Percentage: The percent of runs that
are labeled as “Outside Zone” by PFF for the offense
entering that week on the season. There is evidence
to suggest that play action affects defenses differently
given a team’s scheme (Hasan 2002).
– Inside Zone Percentage: The percent of runs that are
labeled as “Inside Zone” by PFF for the offense entering
that week on the season. There is evidence to suggest
that play action affects defenses differently given a
team’s scheme (Hasan 2002).
– Power/Counter/Man Percentage: The percentage of
runs that are labeled as “Power”, “Counter” or “Man”
by PFF for the offense entering that week on the season.
There is evidence to suggest that play action affects
defenses differently given a team’s scheme (Hasan
2002).
– Down: The down the play occurred on (1st down, 2nd
down, 3rd down or 4th down). Teams run and run
play action at different rates on different downs (Football Outsiders 2019).
– Distance: The yards left for the offense to get a first
down. Teams run and run play action at different rates
depending on the yards needed to obtain a first down
(Football Outsiders 2019).
– Quarter: The quarter of the game the play occurred
in (quarters 1–4 in regulation with overtime labeled as
quarter 5). Game state affects run and play-action rates
(Football Outsiders 2019).
– Seconds Left in Quarter: The amount of time, in seconds, left until the quarter finishes. Game state affects
run and play-action rates. Game state affects run and
play-action rates (Football Outsiders 2019).
– Offensive and Defensive Score Before: The amount
of points the offense and defense have before the play.
Game state affects run and play-action rates. Game state
affects run and play-action rates (Football Outsiders
2019).
– Blitz: Indicator for if the defense blitzed or not. Blitzes
change the composition of coverage behind it.
– Position: PFF’s charted position of a player on that play
based on where they lined up. Player role likely changes
player assignment, so we want to account for that.
– Box Players: The number of players inside the tackle’s
width that are at the first or second level of the defense
to start the play. The composition of the defense changes
the player’s roles.
– Dropback Type and Depth: The type of dropback
(standard, play-action boot, etc.) and how far deep the
dropback went. How far the quarterback drops back
affects how far the linebackers move on the play.
– Coverage Scheme: The category of coverage that
defense ran (Cover 1, Cover 2 man, etc.). Different coverage schemes ask different things of box players.
CORRECTED PROOF
E. Eager and T. Seth: Trade-offs by linebackers using tracking data — 3
– Shotgun, Pistol, Motion: Indicators for whether the
offense lined up in shotgun or pistol or used motion.
Offensive formation and pre-snap motion affect defensive scheme.
– Dome: An indicator for if the roof type of the stadium was a dome. Players move differently in different
environments.
– Turf: An indicator for if the game was occuring on
grass or turf. Players move differently in different
environments.
– Rest: The amount of days the defensive player had
between games. Rest affects player movement and both
team’s scheme due to differences in preparation time.
– Run Direction: The direction of the run (left, middle
or right). Different run directions will affect different
players and their roles differently.
– Rusher Position: The position the rusher is listed as
on the roster (RB, FB, etc.). Different rushers will affect
players and their roles differently.
– RPO: Whether or not the play had a run-pass option
concept as part of it. Different offensive schemes will
affect defensive players and their roles differently.
2.2 NGS tracking data
Next Gen Stats has a series of data points that are collected
through an RFID chip in the player’s shoulder pads that
relays information every tenth of a second. We provide
some justification along with each feature. For the sake
of this paper, we do not include the game-level fatigue of
the player for a couple of reasons, in large part because
it confounds with how good the player is (good players
play more snaps than less good players), and we are trying
to build an expectation that is independent of how good
the player is at football. Additionally, due to the nature of
linebacker play – wherein players are generally facing the
line of scrimmage and stationary at the snap (when they
aren’t blitzing), we did not include speed and orientation at
the snap:
– Relative x Coordinate: Using the ball as the (0, 0)
coordinate, this is how far a player is from the ball
horizontally (sideline to sideline) at the time of snap.
This is a proxy for a player’s role.
– Relative y Coordinate: Using the ball as the (0, 0) coordinate, this was how far a player is from the ball vertically (endzone to endzone) at the time of the snap. This
is a proxy for a player’s role.
– Ball carrier x: The ball carrier’s relative x coordinate at
the time frame. This is a proxy for the offensive player’s
role.
– Ball carrier y: The ball carrier’s relative y coordinate
at the time frame. This is a proxy for the offensive
player’s role.
– Ground Covered: The amount of yards a player covered from their original position at the time of the
snap to 2 s of a run play using the difference Euclidean
distance between the player and the ballcarrier after
2 s. This is the response variable for the E in GCOE.
– Bite Distance: The amount of yards a player moved in
the y direction 2 s into a play-action play. This is the
reponse variable for the E in BDUE. Negative distance
(e.g. displacement) is towards the opponent’s line of
scrimmage (considered bad), while positive distance is
away from the opponent’s line of scrimmage. This is the
response variable for the E in BDUE.
3 Methods
As highlighted in Section 2, we calculated a player’s ground covered
and bite distance on each play to get an expected output based on
the features described above after adjusting for confounders. For both
metrics we used cross validation on a week-by-week level, and then we
used root mean squared error to evaluate the predictions. We explored
a Null Model (taking the mean of the metric), Elastic-Net Regression,
Random Forest, and a tuned XGBoost based on maximum depth and
eta (XGBoost 2022). The XGBoost model was tuned using a hyper grid of
maximum depth between three and six trees with intervals of one tree
and eta between 0.2 and 0.3 with intervals of 0.01, using a five-fold cross
validation technique to minimize the RMSE. Once the best iteration
was found, the model was trained using the best maximum depth and
eta and applied to the data using the week-by-week cross validation
discussed earlier.
3.1 Ground covered over expected (GCOE)
Tackling ball carriers is one of a box player’s primary jobs. Historically,
it was unquestionably their most important jobs, as defenses put an
emphasis on stopping the run. This is likely why play action is so
effective today and is deployed increasingly frequently in both the
National Football League and at the college level (Baldwin 2018a). It also
could be why offenses are using more “run-pass options” (Lee 2021), as
a quarterback can decide which option he wants to deploy depending
on how the box players react to the run fake.
While each of the features in Sections 2.1 and 2.2 influence the
expected ground covered and bite distance for box players, we highlight a few of the relationships between them and response variables
here.
There is evidence to suggest that teams that deploy certain run
concepts (especially outside zone) more frequently are more effective
when they call play action passing concepts (Hasan 2002), and hence
have a different effect on linebacker movement in the running and
play-action game (for a summary of the various run concepts that football teams run, see Palazzolo (2018)). Thus we include the percentage
that a team runs outside zone, inside zone and power/counter/man
run concepts leading up to that game. We also include the particular
CORRECTED PROOF
4 — E. Eager and T. Seth: Trade-offs by linebackers using tracking data
run concept used on a given play, as different run concepts impact
the ground covered by linebackers (see Figure 1). Run plays labeled as
“trick” cause linebackers to flow with the run game the least while the
“man” blocking run concept allows them to flow the most.
Additionally, the PFF charting features include run direction, the
position of the ball carrier, how many box players were deployed
by the defense, whether the run was based off an RPO, along with
indicators for shotgun, pistol and motion. The position of the rusher’s
impact on how much ground was covered by linebackers is evident in
Figure 2.
Lastly, we use the season the play was in, the down and distance,
the quarter, the time left in the quarter, the score of the game, whether
the offense was the home team in the game, and the roof and turf type
of the stadium.
All categorical variables were one-hot encoded. The computer
programming language R (R Core Team 2021) was used to implement
the models. Twenty-two-fold cross validation (for the 18 weeks of the
regular season and four weeks of the playoffs) was used because
of temporal dependencies, where performance in one week may be
influenced by previous weeks. We avoid data leakage by using crossvalidation this way. The results of the different modeling frameworks
are shown in Table 1.
While all of the models were directionally similar, in that the
year-to-year correlations and the season-by-season leaderboards were
roughly the same, XGBoost performed the best, so we used it to create
the expected ground covered predictions.
Figure 3 shows the feature importance of the XGBoost model using
the built in importance feature the R package provides (R Core Team
2021). There were 28 variables used in the modeling process that
became 56 total features due to one hot encoding. The top 20 features
were shown in Figure 3 for simplicity. A player’s x coordinate at the
time of the snap – e.g. how far horizontally from the ball – is the most
important variable, followed by the ball carrier’s y coordinate – how
deep they were in the backfield at the snap. Other positional variables
– both for the linebacker and the ball carrier – are also important
features, as are the run concept both the opposing offense preferred
in the past and ran in the present, which were previewed in Figures 1
and 2.
Once we calculate expected ground covered for each player, we
calculate ground covered over expected (GCOE) by taking the difference between actual ground covered and expected ground covered.
In Section 4.1 we further analyze GCOE for stability and its interplay
with BDUE (Section 4.3). Figure 4 includes a pre- and post-snap example
of two players from a 2017 game and their respective ground covered
estimates.
3.2 Bite distance under expected (BDUE)
Bite Distance Under Expected (BDUE), which was first developed in
Eager et al. (2022) and also studied by Hermsmeyer (2019), has been
updated to adjust for more confounders. Along with rushing the passer,
coverage is the second part of stopping the passing game. We showed
Figure 1: Ground covered by box players as a function of run concept used by the offense. Negative values indicate getting further away from runner
from the start of the snap.
CORRECTED PROOF
E. Eager and T. Seth: Trade-offs by linebackers using tracking data — 5
Figure 2: Ground covered by box players as a function of position of the ball carrier. Negative values indicate getting further away from the runner
from the start of the snap.
Table 1: RMSE values for different proposed modeling frameworks for
GCOE.
Model RMSE
Null model 3.76 yards
Elastic-net regression 3.15 yards
Random forest 2.47 yards
XGBoost tuned 2.47 yards
in Eager and Chahrouri (2019) that coverage performance, as measured by PFF grades, is both highly correlated with team defensive
success in a season and predictive of said success in the following
seasons.
However, one criticism of PFF data is that the coverage grades are
not nearly as stable year-to-year as other grades and metrics. Another
criticism is the grades do not align with conventional opinion – either
media consensus or player earnings data. Some of this instability might
be due to the nature of coverage, as it is dependent on what the opposing offense is doing. By building BDUE, we aim to reduce some of the
noise.
Bite distance is measured as the number of yards, endzone-toendzone, a box player moved from when the ball was snapped to where
they were 2 s into the play. We chose 2 s the measuring point as that
is typically the time it takes the quarterback to fake the handoff to the
rusher and set his feet to find a receiver (although other time thresholds
were tested, the results were qualitatively the same). Similar to GCOE,
we adjust for confounders: the rate at which the opposing offense ran
each run concept coming into the game, the player’s x and y coordinate,
game situation information (down, distance, etc.), whether there was a
blitz (charted by PFF), the player’s position on the play (ILB, FS, etc.),
the quarterback’s dropback depth and type, the coverage scheme and
whether or not it was played in a dome and on turf.
As shown in Figure 5, the position label that a player received
on a play based on their alignment affected their bite distance as
strong safeties usually had more distance of moving up while middle
and inside linebackers usually moved backwards to defend the route
concepts.
The player’s position on a play in the PFF charting data is an
important feature. We only considered players designated as “MLB”
(middle linebacker), “LLB” (left linebacker), “RLB” (right linebacker),
“LILB” (left inside linebacker), “RILB” (right inside linebacker), “SS”
(strong safety), “SSL” (strong safety left) and “SSR” (strong safety right).
Keeping these designations allows us to will keep it broad enough to get
a good sample size on all the players we wanted to analyze, although it
is likely some of these position labels might not always necessarily play
in the box (strong safeties for example).
A quarterback’s dropback depth in yards – as shown in Figure 6
– influenced how much defenders bite as 7, 8 and 9 step dropbacks usually make defenders bite more while 5 and 6 step dropbacks influenced
CORRECTED PROOF
6 — E. Eager and T. Seth: Trade-offs by linebackers using tracking data
Figure 3: Variable importances for final GCOE model. Only the top 20 features (of 28) are shown.
Figure 4: Example ground covered over expected instances from a 2017 game between Kansas City and New England. The figure (a) is pre-snap and
(b) is post snap.
it less. We evaluate multiple models constructed using these features,
using 22-fold cross validation and RMSE, as before. All the models were
directionally similar but the Random Forest model performed the best.
Its feature importance is shown in Table 2.
For expected bite distance, there were 27 variables with 54 total
features due to one hot encoding with the top 20 features being shown
in Figure 7 for simplicity. The “vip” package was used on the random
forest making the scale different from the feature importance Figure 3.
Very similar to the model for expected ground covered, we see features
like the linebacker’s x and y coordinate, the box player’s position and
the offense’s run tendencies pop up as important for determining what
the expected bite distance should be for a player on a play (Figure 7). As
Figure 4 previewed, a quarterback’s dropback depth post-play action
was one of the more important variables as well. Figure 8 includes a
pre- and post-snap example of two players from a 2017 game and their
respective BDUE estimates.
4 Analysis of metrics for individual
players
4.1 Ground covered over expected (GCOE)
One of the primary objectives in developing player-level
metrics, such as GCOE and BDUE, is to create measures
that are more stable and therefore predictive of perfor-
CORRECTED PROOF
E. Eager and T. Seth: Trade-offs by linebackers using tracking data — 7
Figure 5: Bite distance by box players as a function of position they played at the snap of the ball.
Figure 6: Bite distance by box players as a function of the quarterback’s dropback depth on the play-action play.
CORRECTED PROOF
8 — E. Eager and T. Seth: Trade-offs by linebackers using tracking data
Table 2: RMSE values for different proposed modeling frameworks for
BDUE.
Model RMSE
Null model 2.96 yards
Elastic-net regression 2.34 yards
Random forest 2.23 yards
XGBoost tuned 2.27 yards
mance from one season to the next. In contrast to coverage
data, which can be volatile and subject to variability based
on opposing offensive strategies, certain metrics derived
from PFF data, such as player grades and “run stops”,
demonstrate greater year-to-year stability relative to other
American football metrics.
To analyze run-defense-related-metrics, we focus
our attention on players charted as off-ball linebackers
(LLB/RLB, LILB/RILB, or MLB) or strong safeties (SSL/SSR,
SS) by PFF and have recorded 200 or more such run defense
snaps during the 2017–2021 seasons against rushers
classified as running backs or quarterbacks. Our primary
metric of interest is run-stop percentage, which measures
the rate of run stops per run defense snap. We find that
run-stop percentage is correlated year-to-year at a rate of
r=0.41 (Figure 9a), while raw +/−PFF grade (the average of
the play-level grades assigned by PFF graders) is correlated
at a rate of r = 0.46 (Figure 9b).
Despite both of those metrics being good proxies for
run defense stability, GCOE is even more stable with an r
value of r = 0.66 (Figure 10) for the 136 players that played
Figure 7: Variable importances for final BDUE model. Only the top 20 features (of 27) are shown.
Figure 8: Example ground bit distance under expected instances from a 2017 game between Baltimore and New Cincinnati. The figure (a) is pre-snap
and (b) is post snap.
CORRECTED PROOF
E. Eager and T. Seth: Trade-offs by linebackers using tracking data — 9
Figure 9: Year-to-year stability for PFF charting data, with run-stop rate (a) and PFF +/− grade per snap (b). Minimum 200 run-defense snaps in a box
position.
Figure 10: Year-to-year relationship between ground covered over
expected (GCOE) and itself. Minimum 200 run-defense snaps in a box
position.
200 or more run defense snaps in both seasons. GCOE is
also correlated with run stop and PFF grade metrics the
following year, as shown in Table 3.
In a linear model for run stop rate in year n + 1, the
inclusion of run stop rate in year n results in an r-squared
value of 0.17. When GCOE in year n is also included in
the model, the r-squared value increases to 0.20. Similarly,
for PFF +/− grade per run defense snap in year n + 1, the
r-squared value is 0.22 when only PFF +/− grade in year n
is used as a predictor. With the inclusion of GCOE in year n
in the model increases the r-squared value to 0.23.
4.2 Bite distance under expected (BDUE)
Bite Distance Under Expected (BDUE) was already shown in
Eager et al. (2022) to be both a descriptive and predictive
alternative to existing coverage metrics for linebackers,
Table 3: Year-to-year relationship between ground covered over
expected (GCOE) and PFF charting metrics. Minimum 200 run-defense
snaps in a box position.
Metric Correlation Correlation
with GCOE n with GCOE n − 1
Stop rate 0.26 0.33
PFF +/− grade 0.24 0.20
GCOE 1 0.66
which are often unreliable from one year to the next (Figure
11). This is particularly the case in low-sample events,
such as on play-action passes. For instance, pass breakup
rate, which is the frequency with which a player is designated as a primary or secondary coverage player on an
incomplete pass by PFF, has a correlation coefficient of
r = 0.09 from one year to the next for players who have
faced 75 or more play-action snaps within a season (Figure
11a). PFF coverage grades on such plays are correlated at a
an even lower rate (0.04, Figure 11b).
BDUE exhibits greater stability than the aforementioned coverage metrics, as evidenced in Table 4, correlating
at a rate of r = 0.57 year to year (Figure 12). Moreover, BDUE
outperforms PFF +/− coverage grade against play action
in terms of predicting itself from one season to the next.
Specifically, a linear regression model that includes PFF
+/− coverage grade and BDUE in year n yields a regression
coefficient for PFF +/− grade that is not statistically significant at the 0.05 level, indicating that BDUE possesses all
the predictive power. This is an advantageous development
for BDUE, as it addresses a situation where PFF charting
data inadequately captures box players’ performance when
facing opposition play action.
CORRECTED PROOF
10 — E. Eager and T. Seth: Trade-offs by linebackers using tracking data
Figure 11: Year-to-year stabilty for PFF charting data, with pass-breakup rate (a) and PFF +/− grade per play-action snap (b). Minimum 75 play-action
snaps against in a box position.
Table 4: Year-to-year relationship between bite distance under expected
(BDUE) and PFF charting metrics. Minimum 75 play-action snaps faced in
a box position.
Metric Correlation Correlation
with BDUE n with BDUE n − 1
Pass breakup rate 0.02 0.06
PFF +/− grade 0.03 0.15
BDUE 1 0.57
4.3 The relationship between GCOE and
BDUE
In Sections 4.1 and 4.2, we showed that both GCOE and
BDUE improve our understanding of box-player traits by
measuring how well a player flows to the ball on run plays
while also showing how well they see through an offense’s
run fake on a play-action pass. These metrics, moreover,
are related – they illustrate the trade-off between flowing
to the ball in the run game and biting too hard on play
action. Players can either be aggressive in nature (flowing well in the run game but biting hard on play-action)
or conservative (not flowing downhill to the ball on run
plays but not biting on play-action). The correlation between
GCOE and BDUE for players who met the snap thresholds
(n=287) in the same season was−0.40. This makes intuitive
sense, as the more one flows in the run game, the more they
also bite on play-action (Figure 13).
In addition to correlating with each other in a given
season, we see some predictive power for GCOE in BDUE
(Table 5). Other model choices (e.g. elastic-net regressions,
Figure 12: Year-to-year relationship of bite distance under expected
(BDUE). Minimum 75 play-action snaps faced in a box position.
Figure 13: Relationship within year between GCOE and BDUE. Minimum
75 coverage snaps against play action and 200 run defense snaps played
at a box position.
CORRECTED PROOF
E. Eager and T. Seth: Trade-offs by linebackers using tracking data — 11
Table 5: Year-to-year relationship between GCOE and BDUE. Minimum
75 play-action snaps and 200 run-defense snaps played in a box position.
There were 122 players who met both thresholds in adjacent seasons
from 2017 to 2021.
Metric Correlation Correlation
with BDUE n + 1 with GCOE n + 1
BDUE 0.57 −0.40
GCOE −0.40 0.66
and even the null model) produced very similar year-to-year
correlations both within and between BDUE and GCOE.
4.4 Leaderboards
With the stability of our metrics established, we compare
our implied player rankings to consensus opinion. Table 6
displays the league’s best players according to BDUE over
the course of the 2017–2021 seasons.
The list features some of the most accomplished
linebackers of the current NFL era, including several who
are likely to earn a place in the Pro Football Hall of
Fame. Luke Kuechly, for instance, is a strong candidate for
enshrinement, while Dont’a Hightower has a good chance
as well. Jamie Collins, K.J. Wright, Eric Kendricks, and Fred
Warner all secured lucrative contracts based on their outstanding play during their rookie contracts. In addition,
Deone Bucannon, Jordyn Brooks, and Devin White were
all highly regarded prospects and first-round picks in the
NFL Draft, known for their athleticism and talent. White,
in particular, is an intriguing case, as he played a key role
in leading the Tampa Bay Buccaneers’ defense to victory
in the 2020 Super Bowl, despite receiving a low PFF grade.
Table 6: Career BDUE values for linebackers from 2017 to 2021. Minimum
150 play-action snaps played in a box position.
Player Team n BDUE
Dont’a Hightower NE 236 1.21
Thomas Davis Sr. CAR/LAC/WAS 323 0.921
Jamie Collins CLE/NE/DET 316 0.845
Devin White TB 445 0.807
Denzel Perryman LAC/LV 350 0.705
Neville Hewitt MIA/NYJ/HOU 324 0.705
Jordyn Brooks SEA 225 0.678
Luke Kuechly CAR 423 0.663
Eric Kendricks MIN 763 0.625
Deon Bucannon ARZ/NYG/TB 181 0.557
Fred Warner SF 699 0.542
K.J. Wright SEA/LV 371 0.520
This illustrates the difference between this metric and PFF
grades.
Table 7 presents the top-performing players according
to GCOE during the 2017–2021 seasons. This group is not as
distinguished as the leaders in BDUE, which is consistent
with the declining importance of the run game in the modern NFL.
Lastly, taking an average of these two metrics we can
uncover the best all-around linebackers by a combination
of these two measures in Table 8. Even though coverage
is considered a more important kill than stopping the run,
both metrics are weighed evenly for simplicity.
The list comprises some of the top players in the
linebacker position in the league, including Kuechly, Hightower, and Collins, all of whom were previously mentioned. Additionally, there are other players who are highly
Table 7: Career GCOE values for linebackers from 2017 to 2021. Minimum
400 run-defense snaps played in a box position.
Player Team n GCOE
Raekwon McMillian MIA/LV 719 1.01
Zach Cunningham HOU/TEN 1583 0.834
Kiko Alonso MIA/NO 832 0.759
Denzel Perryman LAC/LV 882 0.647
Ja’Whaun Bentley NE 772 0.627
Duke Riley ATL/PHI/MIA 533 0.601
Germaine Pratt CIN 794 0.600
James Burgess CLE/NYG/GB 422 0.590
Sean Lee DAL 609 0.586
Jerome Baker MIA 972 0.570
Kenneth Murray LAC 402 0.566
Jarrad Davis DET/NYJ 1130 0.542
Table 8: Career composite BDUE/GCOE values for linebackers from 2017
to 2021. Minimum 150 play-action and 400 run-defense snaps played in a
box position.
Player Team BDUE GCOE
Dont’a Hightower NE 1.209 0.273
Denzel Perryman LAC/LV 0.705 0.647
Luke Kuechly CAR 0.663 0.268
Sean Lee DAL 0.295 0.586
Jamie Collins CLE/NE/DET 0.845 0.017
Jerome Baker MIA 0.248 0.570
Jordyn Brooks SEA 0.678 0.133
C.J. Mosley BAL/NYJ 0.464 0.319
Zach Brown WAS/PHI 0.380 0.353
Danny Trevathan DEN/CHI 0.427 0.258
K.J. Wright SEA/LV 0.520 0.163
Ja’Whuan Bentley NE 0.013 0.627
CORRECTED PROOF
12 — E. Eager and T. Seth: Trade-offs by linebackers using tracking data
regarded in the league, such as C.J. Mosley, who earned a
high-value free-agent contract from the Jets before the 2019
season. Long-term starters in the league, such as Denzel
Perryman, Sean Lee, and K.J. Wright, are also on the list.
Jerome Baker’s contract with the Miami Dolphins made him
one of the top-10 paid players at the linebacker position at
the time of signing.
5 Analysis of metrics for opposing
teams
In addition to evaluating box players, we also analyze how
opposing offenses influence linebacker movement.
5.1 GCOE against
We examine the offenses for which opposing box players
flow the most and the least against the run, and consider
the maximum GCOE value among box players on a given
run play. While run defense, like pass defense, is a weak-link
system (where outcomes depend on the play of the unit’s
worst players), the linebacker position is more strong link
(depending on the strongest players) than the defensive line,
making this seem like the optimal approach. Table 9 shows
the top five offenses in getting the opposing defense to flow
to the run.
There are some interesting teams in here. Firstly, the
first four teams in the list all have running backs that
have signed top-of-market deals recently, Green Bay’s Aaron
Jones, Carolina’s Christian McCaffrey, Tennessee’s Derrick
Henry and New Orleans’ Alvin Kamara. While there’s a
decent amount of evidence that individual value at the
running back position is less than historical consensus, this
would provide evidence that running backs have value
(Sze Yui 2021). Two LaFleur brothers, Green Bay’s Matt and
New York’s Mike, appear here, giving some credence to the
idea that style of offense causes box players to flow heavily
against run plays.
Table 9: Highest average maximum ground covered by opposing box
players on run plays during the 2021 season.
Season Team Max GCOE against
2021 Green Bay Packers 1.69
2021 Carolina Panthers 1.62
2021 Tennessee Titans 1.62
2021 New Orleans Saints 1.62
2021 New York Jets 1.60
The five best teams in terms of curbing flow from
linebackers on actual running plays are shown in Table 10.
In this list you have three athletic quarterbacks in Kansas
City’s Patrick Mahomes, Jacksonville’s Trevor Lawrence and
Arizona’s Kyler Murray. Linebackers not wanting to be out
of position, and thus being less reactive to “eye candy” like
play action, makes some sense. What’s interesting about
the Steelers being on this list is that their quarterback, Ben
Roethlisberger, was a notoriously awful play-action passer,
as he was the second-worst, worst, and second-worst in
yards per attempt on such plays during his final three years
as a starter, per PFF (PFF). A good question here is a chicken
or egg question; do linebackers move slowly against Pittsburgh because Roethlisberger is not a threat on play action,
or does their movement cause play action not to be not
effective?
The San Francisco 49ers’ offensive play caller, Kyle
Shanahan, has long been lauded for his ability to put
linebackers in conflict, to the point where it’s very likely that
they are not moving as much at all anymore as a precautionary measure. To that point, linebackers against the 49ers
flowed at the seventh-highest rate in Shanahan’s first year
(2017), and eighth-highest in his second year (2018), before
dropping to the sixth-lowest in 2019 and third-lowest in 2020.
5.2 BDUE against
To determine how much or little opposing defenses bite
on play action, we use the average minimum BDUE for
an opposing defense’s box players during the 2021 season. We used the minimum value because the number of
box players on each play varies, making an average value
inappropriate. Furthermore, since defense is considered a
weak-link system, the performance of the weakest link is
crucial to the outcome (Eager 2020). Table 11 displays the top
five teams in 2021 for eliciting high negative BDUE.
Here, we see the teams that employed three of the
top running backs in football in 2021: Tennessee (Derrick
Henry), Cleveland (Nick Chubb) and Indianapolis (NFL rushing leader Jonathan Taylor). Two poor offenses, Detroit and
Houston, round out the list. Neither team had much in the
Table 10: Lowest average maximum ground covered by opposing box
players on run plays during the 2021 season.
Season Team Max GCOE against
2021 Kansas City Chiefs 0.879
2021 Jacksonville Jaguars 0.924
2021 San Francisco 49ers 0.938
2021 Arizona Cardinals 0.971
2021 Pittsburgh Steelers 0.977
CORRECTED PROOF
E. Eager and T. Seth: Trade-offs by linebackers using tracking data — 13
Table 11: Lowest (e.g. most negative) average minimum bite distances
against by box players on play-action passes during the 2021 season.
Season Team Min BDUE against
2021 Tennessee Titans −1.24
2021 Cleveland Browns −1.20
2021 Indianapolis Colts −1.18
2021 Detroit Lions −1.17
2021 Houston Texans −1.13
way of a passing offense in 2021, so defenses could sell out
to stop the run against them, and hence the highly-negative
BDUE values against.
Table 12 looks at the bottom five teams in eliciting
BDUE. Notice that, much like for GCOE, the Kansas City
Chiefs, with one of the best play callers of all time in Andy
Reid and one of the best quarterbacks in all of football in
Table 12: Highest (e.g. least negative) average minimum bite distances
against by box players on play-action passes during the 2021 season.
Season Team Min BDUE against
2021 Washington Football Team −0.436
2021 Kansas City Chiefs −0.569
2021 Buffalo Bills −0.605
2021 Tampa Bay Buccaneers −0.672
2021 Las Vegas Raiders −0.684
Figure 14: Relationship between GCOE and BDUE against for NFL
offenses.
Patrick Mahomes, guide an offense against which linebackers simply don’t move. The Chiefs are among the leaders in both of these for the entire NGS era (2017-present).
Buffalo quarterback Josh Allen is also an athletic quarterback, whose mobility perhaps causing box players to move
less, like what we saw in Table 10. Figure 14 shows the
relationship between GCOE and BDUE against for the 2021
season for the entire NFL.
6 Discussion
In this paper, we build on Eager et al. (2022) to refine
and develop novel metrics for understanding linebacker
instincts in the National Football League. While significant
progress has been made in the evaluation of American football players by researchers both in the public and private
sphere, measures of linebacker play, especially in coverage, have not necessarily aligned with how the players are
valued in the NFL marketplace. Players like Fred Warner,
Jerome Baker, Devin White and others are considered nearreplacement-level or below by grading companies like PFF
at times but are cherished by the league. Some of this
perception is reclaimed in this paper, as the aforementioned players are highly rated by our metrics. Additionally, typical coverage metrics for linebackers haven’t shown
year-to-year stability in the past, but BDUE shows that
there is opportunity for stable linebacker metrics involving
coverage.
Using (x, y) tracking data we have made significant
progress in this work in terms of codifying “traits” who
measurements are usually reserved for the scouting department (Bliss 2022; Eager et al. 2022). A linebacker through
these metrics can earn an “aggressive” label or a “timid” one
based on where they sit in Figure 14. Tracking data provides
a repeatable, consistent and objective means to measure
these traits with data, which significantly aids in the process
of evaluation. Once this tracking data becomes ubiquitous
in the college sphere, projecting linebackers from college to
the NFL will be more sophisticated and robust than in years
past.
Understanding how space is occupied and manipulated
is one of the biggest promises stemming from available
tracking data, as was highlighted in this work. Teams like
the Kansas City Chiefs and San Francisco 49ers, both of
which have made multiple conference championships in
the 2017–2021 seasons, show that their offenses manipulate
offenses well.
Further work in the tracking data sphere should aim
to derive traits for players that are stable and map to
outcomes that we care about in football. For example, in
CORRECTED PROOF
14 — E. Eager and T. Seth: Trade-offs by linebackers using tracking data
Eager et al. (2022) we looked at pass-rush get off rates
in an effort to model athleticism of edge players. This
metric correlated significantly with things like 40-yard
dash, vertical and broad jump. Similar studies for defensive backs would add a great deal to the public understanding of those positions, as up until this point the
state-of-the-industry metrics are not particularly stable or
predictive.
Acknowledgment: The authors would like to thank SumerSports and PFF for their support during this research, and
the Sloan Sports Analytics Conference Research Paper Competition for their feedback on an early version of the BDUE
metric and for funding for Tej Seth’s trip to the conference in 2022. The authors would also like to thank the two
anonymous reviewers, the editors, and associated editors
for their constructive feedback on the initial submission, as
well as Shawn Syed for his comments on the writing of the
manuscript.
Author contributions: All the authors have accepted
responsibility for the entire content of this submitted
manuscript and approved submission.
Research funding: None declared.
Conflict of interest statement: The authors declare no conflicts of interest regarding this article.
References
Baldwin, B. 2018a. Rushing Success and Play-Action Passing. https://www
.footballoutsiders.com/stat-analysis/2018/rushing-success-andplay-action-passing.
Baldwin, B. 2018b. Further Research on Play-Action Passing. https://www
.footballoutsiders.com/stat-analysis/2018/further-research-playaction-passing.
Baldwin, B. 2022. nflfastR, EP, WP, CP, xYAC, and xPass Models. https://www
.opensourcefootball.com/posts/2020-09-28-nflfastr-ep-wp-andcp-models/.
Bliss, T. 2022. Pass Rushing Edge Get Off Speed. https://www.kaggle.com/
code/tombliss/pass-rushing-edge-get-off-speed.
Burke, B. 2013. How Coaches and the NYT Bot Compare. https://www
.nytimes.com/newsgraphics/2013/11/28/fourth-downs/post.html.
Burke, B. 2019. “DeepQB: Deep Learning with Player Tracking to Quantify
Quarterback Decision-Making & Performance.” In 2019 Sloan Sports
Analytics Conference. https://www.sloansportsconference.com/
research-papers/deepqb-deep-learning-with-player-tracking-toquantify-quarterback-decision-making-performance.
Burriel, B., and J. M. Buldu. 2021. The Quest for the Right Pass: Quantifying
Player’s Decision Making. https://statsbomb.com/2021/11/
statsbomb-conference-2021-research-papers/.
Chu, D., M. Reyers, J. Thomason, and L. Y. Wu. 2020. “Route Identification
in the National Football League: An Application of Model-Based
Curve Clustering Using the EM Algorithm.” Journal of Quantitative Analysis
in Sports 16 (2): 121−32,.
Clark, K. 2019. The Race to Make the NFL an Exact Science. https://www
.theringer.com/nfl/2019/4/22/18510099/nfl-draft-analytics-playertracking.
Deshpande, S., and K. Evans. 2020. “Expected Hypothetical Completion
Probability.” Journal of Quantitative Analysis in Sports 16: 85−94.
Eager, E. 2020. How Fragile is Defense: A Look at the Effect of Stars versus
Depth on Game Outcomes. https://www.pff.com/news/nfl-howfragile-is-defense-a-look-ateffect-of-stars-versus-depth-on-gameoutcomes.
Eager, E., and G. Chahrouri. 2019. PFF Data Study: Coverage vs. Pass Rush.
https://www.pff.com/news/pro-pff-data-study-coverage-vs-passrush.
Eager, E., B. Brown, G. Chahrouri, T. Riske, B. Spielberger, L. Sze Yui, Z.
Drapkin, and T. Seth. 2022. “Using Tracking and Charting Data to
Beter Evaluate NFL Players: A Review.” In 2022 MIT Sloan Sports
Analytics Conference. https://www.sloansportsconference.com/
research-papers/using-tracking-and-charting-data-to-betterevaluate-nfl-players-a-review.
Football Outsiders. 2019. Play-Action Passing and Game Conditions. https://
www.footballoutsiders.com/stat-analysis/2019/play-actionpassing-and-game-conditions.
Football Outsiders. 2022. https://footballoutsiders.com.
Hasan, A. 2002. Minnesota Vikings: How Fullback C.J. Ham is the Key to
Understanding the New-Look Offense. https://theathletic.com/
3572885/2022/09/07/minnesota-vikings-cj-ham-offense/?
redirected=1.
Hermsmeyer, J. 2019. Can NFL Coaches Overuse Play-Action? They Haven’t
Yet. https://fivethirtyeight.com/features/can-nfl-coaches-overuseplay-action-they-havent-yet/.
Lee, D. 2021. Lee: A Brief Dialogue About RPOs in the NFL. https://www.pff
.com/news/nfl-dialogue-rpos-miami-dolphins-2021.
Life Magazine. 1969, November 21st Issue.
Lopez, M. 2020. “Bigger Data, Better Questions, and the Return to Fourth
Down Behavior: An Introduction to a Special Issue on Tracking Data
in the National Football League.” Journal of Quantitative Analysis in
Sports 16 (2): 73−9.
Next Gen Stats. 2022. https://nextgenstats.nfl.com/.
Palazzolo, S. 2018. Breaking Down PFF’s Run Concepts. https://www.pff
.com/news/breaking-down-pffs-run-concepts.
PFF. 2022. https:://pff.com.
R Core Team. 2021. R: A Language and Environment for Statistical
Computing. Vienna: R Foundation for Statistical Computing.
Reyers, M., and T. B. Swartz. 2023. “Quarterback Evaluation in the
National Football League Using Tracking Data.” AStA Advances in
Statistical Analysis 107: 327−42,.
Rodrigue, J. 2021. Inside the Rams’ Major Changes to Their Draft Process,
and Why They Won’t Go Back to ’normal’. https://theathletic.c,om/
2507342/2021/04/12/inside-the-rams-major-changes-to-theirdraft-process-and-why-they-wont-go-back-to-normal/.
Sze Yui, L. 2021. Running Backs Don’t Matter: A Dive Back into the
Popular NFL Analytics Phrase. https://www.pff.com/news/nflrunning-backs-dont-matter-a-dive-back-into-the-popular-nflanalytics-phrase.
XGBoost. 2022. https://xgboost.readthedocs.io/en/latest/.