DS502_Final

DS502 Final Project

Outline Submission due Oct 29

https://www.kaggle.com/c/nfl-big-data-bowl-2020

Problem Description

The national football league (NFL) has compiled data on plays where an individual runs the ball. Durring a play like this, an individual carrying the football must run as far as they can down the field with their team's support, while the other team attempts to stop them. The NFL wishes to predict, in real time, how many yards the player with the ball will run. This prediction will be used by the NFLs analysts durring and after plays throughout the football season to better understand what contributes to more successful runs.

Data Description

The following exerpt from the NFL Big Data Bowl Kaggle page describes the data available:

Each row in the file corresponds to a single player's involvement in a single play. The dataset was intentionally joined (i.e. denormalized) to make the API simple. All the columns are contained in one large dataframe which is grouped and provided by PlayId.

GameId - a unique game identifier PlayId - a unique play identifier Team - home or away X - player position along the long axis of the field. See figure below. Y - player position along the short axis of the field. See figure below. S - speed in yards/second A - acceleration in yards/second^2 Dis - distance traveled from prior time point, in yards Orientation - orientation of player (deg) Dir - angle of player motion (deg) NflId - a unique identifier of the player DisplayName - player's name JerseyNumber - jersey number Season - year of the season YardLine - the yard line of the line of scrimmage Quarter - game quarter (1-5, 5 == overtime) GameClock - time on the game clock PossessionTeam - team with possession Down - the down (1-4) Distance - yards needed for a first down FieldPosition - which side of the field the play is happening on HomeScoreBeforePlay - home team score before play started VisitorScoreBeforePlay - visitor team score before play started NflIdRusher - the NflId of the rushing player OffenseFormation - offense formation OffensePersonnel - offensive team positional grouping DefendersInTheBox - number of defenders lined up near the line of scrimmage, spanning the width of the offensive line DefensePersonnel - defensive team positional grouping PlayDirection - direction the play is headed TimeHandoff - UTC time of the handoff TimeSnap - UTC time of the snap Yards - the yardage gained on the play (you are predicting this) PlayerHeight - player height (ft-in) PlayerWeight - player weight (lbs) PlayerBirthDate - birth date (mm/dd/yyyy) PlayerCollegeName - where the player attended college Position - the player's position (the specific role on the field that they typically play) HomeTeamAbbr - home team abbreviation VisitorTeamAbbr - visitor team abbreviation Week - week into the season Stadium - stadium where the game is being played Location - city where the game is being player StadiumType - description of the stadium environment Turf - description of the field surface GameWeather - description of the game weather Temperature - temperature (deg F) Humidity - humidity WindSpeed - wind speed in miles/hour WindDirection - wind direction

Prediction type

The problem could be approached as either regression or classification. We will approach the problem as a regression problem, for a couple of reasons. First, there would be a large number of classes with little data in each class if we approached it as a classification problem. Second, a classification problem would predict only integer values for a single play's yardage, while a regression can incorporate continuous values. While we will need to submit a file which predicts yards as discrete values, using a regression model will still allow for more precise measurements for our predictions because it allows for the fact that yardage is not discrete in real life.

Methods

Before we begin constructing the model we'll need to attempt to convert the categorical atributes to ordinal atributes. For example, HomeTeamAbbr could be converted to information on the offesive and devensive power of the team. Other categorical atributes can be one hot encoded. There is a possibility of developing a simulation of the play which doens't use machine learning. The advantage of this is that it's easier for a machine learning model to learn the difference between what a theory-based model predicted and the truth than it is for a neural network to learn the theory itself. The disadvantage of this theory-based model creation is that we might not have enough time to work on the machine learning if we use it all making a theory-based model. After we have finished processing the data we plan on using a neural network. The first neural network we'll attempt to use is a regular feed forward network. Although unlikely, it is possible that at any given point in a play, we have enough information to know how many yards the ball carrier will travel. It is likely though that the previous states of the play are relevant when predicting the total yards run, therefore the second model we will construct is an RNN with LSTM nodes. LSTM nodes are better than normal perceptrons because they can learn for themselves how long past information stays relevant. An alternative to the RNN is a 1-D CNN. The 1-D CNN will look for trends in a stream of play data, which may lead to other insight useful to predicting the total yards run that an RNN might not have been able to see. It is likely that multiple models will perform well, and if so, an ensembling process will be used to make the final prediction.

We can do some neural networks, but the point of our analysis can be how different models perform and why. We could say like, 4 tweaks on model X, or just 4 models

Error Metrics and Algorithms for assessing them

The kaggle competition stipulates use of the Continuous Ranked Probability Score (CRPS) based on the cumulative distribution of yards gained. The formula is as follows: $$C = \frac{(1)(199N)} \sum{m=1}^N \sum{n=-99}^99 i^2 (P(Y \leq n) - H(n - Y_m)^2)$$ where P is the predicted distribution, N is the number of plays in the test set, Y is the actual yardage and H(x) is the Heaviside step function. The kaggle competition will not score the submission if the CDF is a decreasing function.

Comments and Concerns

The NFL data has an interesting structure which gives us unique insights and challenges. Rather than each row being an observation, collections of rows unified by a 'Play I.D.' represent one play during a football game, with a response yards gained. There are many interesting preprocessing techniques to transform this data into something we can model. There are countless exploration topics in the data on topics such as matchups, player performance, and team organization that can motivate our models and insights about this unique data.

The high variability in the data is one concern for our regression. An identical play (players, stadium, latent factors) could lead to a 0 yard gain or a 100-yard touchdown. For this reason we will have to negotiate with extreme noise in the developement of our model, and in some cases a lack of predictor correlation to the response.