Neural Network Prediction of NFL Football GamesNFL Prediction via Neural Network.pdfTable of ContentsIntroductionWork PerformedData CollectionData ExtractionPreliminary StudyTraining and Prediction Set CreationNeural Network SelectionNeural Network Parameter SelectionData PreprocessingMaking PredictionsResultsBaseline StudyDiscussionReferencesAppendix 1 – Sample Box ScoreChicago Bears versus Green Bay PackersWeek 14 – December 7, 2003Appendix 2 – Source CodeNeural Network Prediction of NFL Football Games Joshua Kahn ECE539 – Fall 2003 December 19, 2003Neural Network Prediction of NFL Football Games Table of Contents Introduction......................................................................................................................... 2 Work Performed..................................................................................................................2 Data Collection ............................................................................................................... 3 Data Extraction ............................................................................................................... 3 Preliminary Study ........................................................................................................... 3 Training and Prediction Set Creation.............................................................................. 5 Neural Network Selection............................................................................................... 6 Neural Network Parameter Selection.............................................................................. 6 Data Preprocessing.......................................................................................................... 8 Making Predictions ......................................................................................................... 8 Results................................................................................................................................. 8 Baseline Study ................................................................................................................ 9 Discussion......................................................................................................................... 10 References......................................................................................................................... 11 Appendix 1 – Sample Box Score...................................................................................... 12 Appendix 2 – Source Code ............................................................................................... 13 Page 1 of 19Neural Network Prediction of NFL Football Games Introduction Over the past decade, football has truly become America’s game. Now, with millions of people watching from their easy chairs every Sunday, the National Football League has become a multi-billion dollar business. There are countless internet sites that claim that they “know” the outcomes of future games. These sites come in many varieties, from the trustworthy (such as ESPN.com) to the downright seedy (just do a search for NFL football betting). With a plethora of different opinions out there, the question becomes, which ones are actually correct? Most of the available prognostications are based on human opinion, which invariably leads to some degree of bias. During this project, a completely objective system was designed to predict the outcome of future NFL games, purely for academic purposes, of course. The problem in designing such a system, however, is the high number of “intangible” aspects in every game. For example, towards the end of the season a team with a losing record may play harder to save a favorite coach from being fired. Such a scenario would be difficult to predict from a purely objective standpoint. In general, however, it would seem plausible that it is possible to predict the outcome of a game based only on the statistics of the competitors. The NFL keeps a very large set of statistics for each game played, much beyond those found in the box scores of a typical newspaper. Many prognosticators study “splits,” or statistics broken down into extremely specific categories. For instance, data exists on the performance of every skill position player (quarterback, running back, wide receiver) when they play in cold versus warm weather, indoors versus outdoors, and even their performance on a given down with a given amount of yardage necessary to get a first down. The amount of available data becomes overwhelming when this information is considered; therefore, some small subset of data must be chosen that accurately predicts the outcome. This, of course, is easier said than done. For this project, it was determined that the data used in making a prediction would be available from a standard box score, thus cutting down on the amount of available data substantially. Still, box scores provide a great deal of information on each game and the importance or relevance of each statistic must be determined prior to making a prediction (this will be discussed further in the next section). A final consideration here is that winning occurs in a variety of ways. A quick glance at the total yardage row in a box score does not correctly identify the winner of the game, although it does lend some insight. In fact, there are countless examples of games in which the team with more total yardage actually lost. This indicates that there is no linear mapping method in which a winner can be chosen based solely on a group of statistics. Instead, a neural network could be used to perform non-linear mapping based on a variety of statistics. Work Performed In order to build a neural network that was capable of accurately predicting the outcome of a NFL football game, several steps were required. Page 2 of 19Neural Network Prediction of NFL Football Games Data Collection As mentioned earlier, it was determined early on that all attempts would be made to ensure that the statistics used for prediction would be available from a typical NFL box score. Also, as previously mentioned, winning occurs in a variety of ways. An easily visible pattern does not exist that indicates which team will win (if it did, this problem would be easily solved); therefore, a large amount of data must be provided. A large data set would include a variety of modes that win
View Full Document
Unlocking...