Inspiration
The math works. Over the course of a season, there's some predictability to baseball. When you play 162 games, you eliminate a lot of random outcomes. There's so much data that you can predict: individual players' performances and also the odds that certain strategies will pay off. -Billy Beane
In 2002, the Athletics became the first team in the 100+ years of American League baseball to win 20 consecutive games by applying sabermetric methods. There are many models for different games but I didnt come across any such paper or article which tells us the method to create a model for soccer transfer market recommendation.
What it does
It aims to provide transfer market recommendations based on the player's potential and recent performances that would help the team perform better (with given constraints).
Step-wise Method
a. Predict the market value of the player based on input feature b. Predict the Form of the player based on his last 5 year performance c. Derive an equation for Impact Score of the player (combination of player potential and recent history) d. Calculate the MVP of the player which is a function of Market Value and Impact Score e. Use constrained programming to come up with a set of optimal recommendations for the transfer market.
How I built it
I used the following data:
FIFA 18 All Player Database Fantasy Points and Player Statistics for Top 5 Leagues in Europe Scraping Websites for League Performance Data Manually Searching and Gathering Transfer Budget for Clubs
- Predicting the Market Value of the Player
Supervised learning using Ridge Regression (alpha = 0.01) and Feed-forward Neural Network (1024 * 1024 * 1024 hidden layers) Input Feature - 75+ input features like FIFA attributes, Transfer Release Clause, Age, Characteristics on and off fields RMSE - 696431 (value in millions - upto 100 million) Score - almost perfect 1.0 (17994 data-points, 75+ features) - NN 0.9597 - Ridge with alpha = 0.001
- Predicting Form of Past Seasons
Supervised learning using Ridge Regression and Feed-forward Neural Network (1024 * 1024 * 1024 hidden layers) Input Feature - 18 input features like Goals, Assists, Cards, Fantasy Points, Appearances, etc Score - Very low as we don’t have good features(3923 data points - 18 features and variability in actual performance, not so easily modelled.
- Calculating the Impact Score
Evaluate different methods to calculate the form of the player based on the forms of the past 5 years form. (Weight Decay) Semi Supervised Learning to infer the “Impact Score” of the player based on the different weights on Overall FIFA score and Form obtained from the above model.
- Team’s Last Season Performance (for Top 5 Leagues)
Analyse the area of weakness (defence, attack) by normalizing Goal Scored, Goal Conceded, W/D/L ratio with respect to rest of the teams.
- Constrained Programming (NOT IMPLEMENTED)
Find multiple solutions of optimal team by replacing as few players as possible while maximising the impact score. Constraints - Budget (Market Value + Weekly Wage), Squad Size, Player Position Challenges - Trade-off between quality and quantity Trade-off between positions (attacker vs defender) Brute-Force is not practical
Challenges I ran into
a. DATA (more than 90% time was spent on data collection) b. 1-1 Mapping between different datasets (Manual annotation) - ~ 2 hrs to run once c. Important Data points missing for calculating impact score (match-winning goals/assists, penalty wins, accuracy etc) d. Incomplete Data in FIFA (not all clubs in Top 5 are listed) e. No PRIOR Work for Model Creation and lot of constraints f. Manually creating different functions for Impact Score.
Accomplishments that I'm proud of
a. Understanding and defining a challenging problem statement b. Understanding the model behaviour and the scope for improvements c. Attemting to do something not done before as an individual (No Team)
What I learned
Its very hard to model real-life events to a mathematical model. Collecting data is one of the most challenging tasks of machine learning. I learned to experience the effects of regularization, bias-variance trade-off on real data-sets and see how to tune different hyperpaprameters.
What's next for Recommendation System for Soccer Transfer Market
a. Getting the time-constraint working to get the optimal solutions b. Collecting more data points to improve our Form Accuracy c. NLP on Football Commentary to infer important events of a match d. Applying Game Theory (Online Learning Theory) to implement strategies against specific teams. e. Optimized-Squad Selection and strategies based on Frame Analysis and heat-maps.
Log in or sign up for Devpost to join the conversation.