Introduction to Data Analysis for predicting player performances in sports

At this webinar, we will talk about applications of Data Science methods to the sports analysis, using prediction of the goals scored by a player in a football game as a sample. We will cover some fundamental aspects of the domain, explain what features w

Transcript

1. Introduction to Data Analysis for predicting player performances in sports Alexandru Gris, Mykhailo Kobernyk, Scout Gaming Tech
2. Player performances in football Usually, player-specific events in football (goals, assists, yellow/red cards, own goals etc) are happening up to 10 times per game All events we’re going to consider today are distributed by Poisson distribution law: P(X = k) = λke−λ k!
3. Poisson regression We can use Poisson regression when dependent variable is under the Poisson distribution Independent variables can be either numerical or categorical It’s part of the General Linear Regression (GLR) models family Formula: log(λ) = β0 + β1 x1 + β2 x2 + . . . + βn xn
4. Data normalization In some cases (sum of rare events on a relatively small amount of time), our independent variables contain skewed data We’re normalizing data before further processing to prevent biases in our linear model
5. Let’s switch to the first practical task
6. Bayesian methods Instead of an attempt to find the frequency of some event (e.g., goal of some specific player), we’re going to interpret expectation of the event in future based on the observations in past and our understanding of the event distribution In python, Bayesian inference can be implemented either with pymc library or TensorFlow Probability The most common methods are Monte Carlo simulation and Markov chain- based sampling algorithms (e.g. Metropolis)
7. Fantasy sports User is allowed with some amount of virtual “budget” that can be used to “buy” players into his/her team. The stronger the player — the more expensive he is. During the real life matches, players are scoring goals, assists or other events, and each of them is priced with some amount of “fantasy points” for the user’s team. User’s goal is to pick the players who are going to score the most “fantasy points” at total. At least, more then other user’s picks. :-)
8. Fantasy duels Using the same scoring rules, we can make a “duels” between real players. Player who scored more fantasy points during the game is considered a winner. In our case, we have “odds” for each player. Smaller odds correspond do the higher player’s probability to win the game. In case of the draw duels is cancelled.
9. Further readings and links Cameron Davidson-Pilon “Bayesian Methods for Hackers” Robert Kissell, Jim Poserina “Optimal Sports: Math, Statistics, and Fantasy” https://stats.idre.ucla.edu/stata/dae/poisson-regression/ https://medium.com/statistical-guess/z-test-z-statistics-e2dd1782656d https://en.wikipedia.org/wiki/Metropolis-Hastings_algorithm https://www.techradar.com/news/what-are-daily-fantasy-sports Notebook 1 at https://tinyurl.com/y8bajlox Notebook 2 at https://tinyurl.com/ydxpcjea