class: center, middle, inverse, title-slide # Run vs. Pass Play Prediction: ## Incorporating NFL Tracking Data ### Tej Seth ### Nicole Tucker ### November 6th, 2021 --- # Previous Play Type Prediction Models- NFLFastR - Situational Factors from Play by Play Data - Quarter - Down - Yards to Go - Yardline Number - Half Seconds Remaining - Score Differential - Shotgun Formation - XGBoost Model - Accuracy: 70.1 % --- # The Data NFL 2017 Tracking Data Week 1 Through Week 6 Corresponding NFL Play by Play, Games and Players Data NFL FastR 2017 Play by Play Data
nflId
x
y
dir
dis
s
event
displayName
frame.id
playId
gameId
2495340
41.56
16.54
78.90
0.41
3.91
NA
Anthony Sherman
1
44
2017090700
2495340
41.95
16.62
79.16
0.40
4.28
NA
Anthony Sherman
2
44
2017090700
2495340
42.40
16.73
79.46
0.47
4.66
NA
Anthony Sherman
3
44
2017090700
2495340
42.85
16.82
79.76
0.46
5.04
NA
Anthony Sherman
4
44
2017090700
2495340
43.36
16.92
80.12
0.51
5.39
kickoff
Anthony Sherman
5
44
2017090700
--- # The Data <img src="BDB_Schema.png" width="70%" style="display: block; margin: auto;" /> --- # Created a Baseline Situational Factors Predictive Model for Comparison Purposes - Predictors: - Quarter - Down - Yards to Go - Yardline Number - Half Seconds Remaining - Score Differential -- - Accuracy: 63.27% - Brier Score: 0.2085 --- # Defining Variables Created from Tracking Data - Width of Formation + Standard Deviation - Width of Offensive Line + Standard Deviation - Depth of Offensive Line <img src="olinediagram.jpeg" width="50%" /> --- # Defining Variables Created from Tracking Data - Width of Formation + Standard Deviation - Width of Offensive Line + Standard Deviation - Depth of Offensive Line - Deep Backfield RB- Indicator Variable - QB Position (Under Center, Shotgun, Pistol) - Personnel (Number of WRs, RBs, TEs) - Fullback- Indicator Variable - Maximum Wide Receiver Distance Off Line of Scrimmage - Man in Motion- Indicator Variable - Tight End Starting Near Ball- Indicator Variable <img src="setdiagram.jpeg" width="60%" /> --- ## A Larger Formation Width May Indicate a Pass Play <img src="width.png" width="65%" style="display: block; margin: auto;" /> --- ## A Larger Linemen Width Standard Deviation May Indicate a Run Play <img src="linemenwidthse.PNG" width="65%" style="display: block; margin: auto;" /> --- ## Using a Larger Sample Size to Model Expected Pass for Situational Factors - Trained model on 2012-2017 NFLFastR Play by Play Situational Data - Created a Situational Expected Pass Variable <img src="DownYardlinePROE.PNG" width="55%" style="display: block; margin: auto;" /> --- ## Comparing Model Types <img src="CompareChart.jpeg" width="75%" style="display: block; margin: auto;" /> --- ## Using a Mixture of Tracking Data Variables and Situational Variables Produces the Most Accurate Model - Situational Factors and Tracking Data Variables in a Random Forest Accuracy: 75.2 % - Situational Factors and Tracking Data Variables in a Random Forest Brier Score: 0.167 <img src="VarImpt.PNG" width="55%" style="display: block; margin: auto;" /> --- ## The Predictability of Play Type Varies by Team <img src="Predictability.PNG" width="65%" style="display: block; margin: auto;" /> --- ## Each Team's Pass Rate Over Expected for 2017- Accuracy <img src="PROE.PNG" width="65%" style="display: block; margin: auto;" /> --- ## Each Team's Pass Rate Over Expected for 2017- Brier Score <img src="xpBrier.PNG" width="65%" style="display: block; margin: auto;" /> --- ## Mixed Effects Model- Effect of Team <img src="mixedeffects.PNG" width="65%" style="display: block; margin: auto;" /> --- ## Model Strengths- Down and Distance <img src="BrierDownDist.PNG" width="65%" style="display: block; margin: auto;" /> --- ## Model Calibration- Field Position <img src="yardlinecalibration.PNG" width="65%" style="display: block; margin: auto;" /> --- ## Model Calibration- Down and Distance Situation <img src="downdistcalibration.PNG" width="65%" style="display: block; margin: auto;" /> --- ## Interpreting the Model <img src="coefficients.JPG" width="65%" style="display: block; margin: auto;" /> --- ## Expanding Our Model to Update Probability of Pass After Snap Updates every 0.10 of a second up to 2.5 seconds after the snap -- Variables Created: - Average Offensive Linemen Distance from Line of Scrimmage - Average Offensive Linemen Speed - Average Receiver Distance from Line of Scrimmage - Average Receiver Speed - Probability of Pass Before Snap --- ## Average Prediction Accuracy Every 1/10 Second <img src="avgaccovertime.jpeg" width="65%" style="display: block; margin: auto;" /> --- ## Single Play Probabilities <img src="final-brady.gif" width="70%" style="display: block; margin: auto;" /> --- # Completed Work - Created variables using tracking data - Created a basic model using only play by play data - Created models including tracking data variables - Investigated the predictability of each team - Looked at how our model does in specific subsets - Expanded our model to update predictions within the first 2.5 seconds of the play # Acknowledgements A big thank you to our project advisors! - Brian Macdonald - Sam Ventura And a big thank you to our program instructor, Ron Yurko, for all the help and support this summer!