I created this website to host some of the projects that I have created. Here is a description of each one:
Expected Points Per Drive - When creating a model to predict which team will win NFL football games, the final scores should be normalized for the amount of opportunities teams had to score in each game. I created a script in R that properly assigns the EPA (https://www.nfeloapp.com/analysis/expected-points-added-epa-nfl/) from nflfastR to each of the 8 units that make up a football team: Offense, Defense, Kickoff Kicking, Kickoff, Receiving, Punt Kicking, Punt Receiving, Field Goal/Extra Point Kicking, Field Goal/Extra Point Defending. I made sure that, for every game, the sum of the values assigned to each unit equals the actual score differential. This required reporting multiple bugs to nflfastR that were quickly fixed. By dividing the value assigned to each unit by the number of drives in that game, I created a descriptive metric that normalizes for game length while also being perfectly correlated with the actual game outcome. This metric (Expected Points Added per Drive or EPApD) can be used as a target variable for models to predict future points per drive of each unit and therefore game outcomes. I host a CSV file of this data for every game back to 2011. I wrote an article explaining the theory and practice for creating this metric. I also included graphs showing the average EPApD's for each unit in the 2021 regular season.
Maximizing Expected Growth - I derived an equation that maximizes the expected growth of seed money given betting lines and the actual probabilities of the events. I explain the theory and application of the formula in an article using visualizations.
Clustering - I used Bayesian reasoning to create an algorithm that inputs user data about pairs of items and then outputs a clustering for the items. The data is whether the user thinks random pairs of items are Similar, Not Similar, or Completely Different. The data entry is facilitated through the website and then processed in a Google Colab notebook that also explains the algorithm and all of the theory behind it. In the future I will add the ability to concurrently create an overall ranking of the items while getting data for clustering. Instead of the pairs of items being random, they will be chosen to create the overall ranking in as few comparisons as possible.
Dynasty Football Trade Calculator - I am creating a trade value caluclator for dynasty fantasy football using actual trade data from Sleeper. I include a page that explains the theory behind the calculations.
Fantasy Football Target Variables - The value of a fantasy football player comes from their ability to win you the championship. This value comes from the amount of yards, touchdowns, etc. that they gain each week, the number of games they play, and the number of games you start them. I transform all of these data points into their uncorrelated aspects that can each be used as target variables when building models to predict future fantasy value. I also supply the transformation that takes the predicted amount of production and translates it into fantasy points, then likelihood of winning a game, then likelihood of winning the championship based on scoring settings, roster requirements, number of teams, and playoff structure.
Conquer Wordle - I explain how I optimized Alpha-beta pruning to find the optimal words for two different metrics of Wordle play: minimizing the maximium number of guesses and maximizing the percent of words that take three guesses. I then link to a Google Colab sheet with these algorithms.
Future Projects:
- Use a Markov decision process to build a new version of EPA from models that predict state transitions instead of the next scoring play.
- Predict winners of NFL games using a Bayesian Network that takes in past outcomes and metrics derived from nflfastR data. I will also potentially add mertics derived from PFF grades and other prermium data.
- Build models to predict distributions for 'Fantasy Football Target Variables' for redraft and dynasty.