Every year in Major League Baseball (MLB), hundreds of players become free agents. During this time, their agents are consistently meeting with both their clients (players) and the general managers who they are negotiating these contracts with. Sometimes, players can avoid free agency altogether by signing a contract extension with their current team prior to their contract expiring.
Over the past few years, there have been situations where players have too high of expectations for the contract they will receive, due to bad advice and overly optimistic thinking by their agents. In March of 2021, the New York Mets reportedly offered their outfielder Michael Conforto, a 5 or 6 year contract worth $100-120 million. He declined this contract after consulting with his agent, Scott Boras. A year later Michael Conforto remains unsigned 2 months into the 2022 MLB season and has undergone season ending surgery on his shoulder, meaning he will earn $0 this season. During the offseason before the 2018 season, infielder Mike Moustakas was offered a qualifying offer, which is a one year contract worth $17.4 million. After consulting with his agent, Scott Boras, he declined this offer seeking a multi year deal. Moustakas ended up signing a one year, $6.5 million contract that offseason. The next offseason, he signed a one year, $10 million contract. If he had been given proper advice, he would have made more money in the 2018 season than he ended up making in the 2018 and 2019 seasons combined.
Agents not having realistic expectations and not understanding what their clients will be offered by general managers, puts them at a competitive disadvantage. It can make their clients upset, and can lead to being fired, when the players don't receive the contract offers they were told by their agent to expect. It also puts agents at a competitive disadvantage in negotiations with general managers.
I was hired by the Boras Corporation to create a predictive model to better predict MLB free agent hitters salaries. After these major blunders with Michael Conforto and Mike Moustakas's contracts, the Boras Corparation has realized that they need to be able to better predict free agent salaries so they can tell their clients what to expect, and to give them a leg up in negotiations with MLB general managers.
In order to address the problem that the Boras Corporation is facing, I will use the following regression models to improve the companies ability to predict free agent salaries:
- Random Forest Regression
- XGBoost Regressor
- XGBoost Random Forest (XGBRF) Regressor
I chose to use these models because they all can handle multicollinearity and have the ability to tell me feature importance. Feature importance is important because the agents can tell their clients what part of their game to improve during the offseason to give them the best chance at a bigger contract.
My original dataset was found at Kaggle.com. The dataset contained 20 CSV files that included MLB data from 1871-2021. Of those 20 CSV's I used the following:
- AllstarFull.csv
- AwardsPlayes.csv
- AwardsSharePlayers.csv
- Batting.csv
- People.csv
- Salaries.csv
- Teams.csv
While the batting.csv statistics went through 2021, the salaries only went through 2016, so I wanted to find data for the salaries from 2017-2021. I found this data at Cot's Baseball Contracts. This data included salary data from 2000-2021, and helped me fill in 1600 values in the original salary data that were null. While searching for more salary data to fill in the null values, I found data at The Baseball Cube. This data included salary data from 1985-2022. I purchased this data set for $20. It helped me fill in 6,449 null values in the original salary data.
I subsetted my data to only use data from 1984-2021. I also subsetted my data to only include players who played at least 6 seasons (so the only players included were ones who played long enough to make it to free agency). I then made a new dataframe where the columns of the dataframe were the players' hitting statistics for each of his first 6 seasons, and included the players' average salary of his 7th, 8th, 9th, and 10th seasons (or less if he didn't play 10 seasons). This created a dataframe that was 1063 rows (one for each player in my dataset that played 6 or more seasons) by 131 columns. The columns included the players' hitting statistics for his first 6 seasons, along with whether he made the All Star Game, whether he started the All Star Game, whether he won an award, or got votes for MVP or Rookie of the Year, whether his team made the playoffs, and if his team won the World Series.
I subsetted my data to only use the data from 1984-2021 and to only include players who played at least 6 seasons (so the only players included were ones who played long enough to make it to free agency). I then made a new dataframe where the columns of the dataframe were the players' hitting statistics for each of his first 6 seasons, and included the players' average salary of his 7th, 8th, 9th, and 10th seasons.
I used Regression Models that were not affected by multicollinearity because most baseball statistics are correlated to each other. I used Random Forest, XGBoost, and XGBoost Random Forest and I scored my models using R-squared and RMSE. Using RMSE allowed me to see how far off my model’s salary predictions were in terms of dollars.
My top 3 models scored pretty similarly. My best model was one of my Random Forest gridsearch models. This model had an R squared score of .68944, which means my model was able to account for nearly 69% of the variance within my data. My final model has an RMSE of just over 3,000,000, which means the error of our salary predictions is just over $3 million. This seems like a fair amount of error to expect when agents and general managers are negotiating a contract. However, I believe that once I am able to detrend the salary data, my model will have an even lower RMSE.
My final model was a Random Forest gridsearch. This model is able to account for nearly 70% of the variance within my data. My final model has an RMSE of 3,063,911, which means the error of our predictions is just over $3 million. This seems like a fair amount of error to expect when agents and general managers are negotiating a contract. However, I believe that once I am able to detrend the salary data, my model will have an even lower mean squared error. This is important because the average salary in MLB has gone from $371,571 in 1985 to $4,414,184 in 2022, according to Baseball Reference. By using regression models to help predict MLB free agent salaries, I have given The Boras Corporation a competitive advantage when dealing with clients and when negotiating new contracts with General Managers.
- The fielding data that was included was very incomplete and did not include any of the truly relevant fielding statistics.
- The original salary data was pretty incomplete, and without the data set purchased from The Baseball Cube it would have been very difficult to succesfully complete this model, because I would not have had enough data.
- Find a way to account for the 2020 Pandemic Shortened 60 Game Season.
- Detrend the salary data so the yearID isn't the most important feature and the players' statistics were more important.
- Include data on players' service time so I can more accurately predict when a player is going to free agency.
- Find reliable fielding data so I can include a players' fielding statistics.
Please review my Data Cleaning Notebook, Modeling Notebook, or Presentation Deck.
For additional questions, feel free to contact me.
├── Data <- Source data .csv files
├── Images <- Exported Notebook visualizations and Notebook Images
├── Workspace <- Additional scratch work notebooks, and CSV files for my app demo
├── Data_Cleaning.ipynb <- Data Cleaning and EDA in Jupyter Notebook
├── Modeling.ipynb <- Preprocessing and Modeling in Jupyter Notebook
├── MLB_FA_presentation.pdf <- PDF version of project presentation
├── app.py <- Code for Streamlit application
├── environment.yml <- My environment for you to recreate my project
└── README <- Top-level README for reviewers of this project
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent