In the National Football League (NFL), a team’s general manager (GM) is tasked with building the best possible roster. The GM can acquire players through free agency (a ‘free agent’ is a player who is currently not contracted to any team), the draft (an annual process by which teams ‘select’ new players into the league, usually college players), or trades, also having to account for constraints such as roster size and salary cap (the maximum a team can spend in a year). Our goal with this project was to take on the role of a GM through data science lens. We leveraged Python’s BeautifulSoup library to collect historical data for the NFL, including salary, draft, Approximate Value (AV), roster, draft combine, and standings data. We then divided our project into three parts: player and draft pick evaluation (Statistics), draft prospect k-nearest neighbor predictions (ML), and team composition analysis (ML). For player and draft pick evaluation, we used Pro Football Reference’s AV metric to create a ‘fair salary’ metric for a player. We combined historical draft AV data with this metric to arrive a ‘fair salary’ for draft pick to value future draft picks. By comparing the ‘fair salary’ of draft picks to their actual predetermined salary, we were able to determine which draft picks offered the most purchasing power given salary cap restraints. Next, we implemented a k-nearest-neighbors test to determine which historical players this year’s 2017 NFL Draft prospects matched most similarly to, based on NFL combine test results only. Finally, we explored optimal team composition (across positions), using position-wise summary statistics for team rosters as feature sets to predict the playoff performance of real and hypothetical teams. Our results for the player and draft pick evaluation confirmed that draft picks tend to serve as cheap talent, but offered the caveat that top draft picks may in fact be overvalued. Our draft prospect k-nearest-neighbor test was an exploratory aside, and offers value in terms of player comparison, as well as generally confirming that teams look at more than physical traits when evaluating players. Finally, our team composition analysis lent credence to the ideology that “defense wins championships,” while also championing the idea that quarterbacks are crucial to team success. Limitations of our study include lack of data (the NFL only plays 16 games a year, and there is wide variance in game outcomes), as well as the fact that there are practically infinite variables that affect a team and player’s success that we couldn’t entirely account for. Despite this, our results appear useful in controlled settings (player comparison on physical traits), and in some cases provide promising ground for continued exploration (draft pick evaluation, optimal team composition).
jordanholbrook / nfl-player-valuation Goto Github PK
View Code? Open in Web Editor NEWThis project forked from kevinli96/nfl-player-valuation
Analysis of NFL Player Valuation
Home Page: https://kevinli96.github.io/NFL-Player-Valuation