Maximizing Roster Efficiency in the MLS

Oliver Daboo, Otis Birnbaum

Colby College, Williams College

External Advisor: Daniel Wicker, Charlotte FC

1

2

Motivation

• MLS is the only major professional soccer league in the world with a salary cap and has a lot parity.
• Lower budget teams have a chance to compete if they find players who contribute a lot relative to their wage

​

Main Question

• How should MLS teams construct their rosters and salaries spreads based on a players position, age, and nationality?

Data

• MLS Seasons 2021-2024 from American Soccer Analysis
• Used xG Difference to evaluate team performance
• Used Goals Added, a value added metric, to evaluate player performance
• Created a goals added per 90 minutes played per $10k (ga_per_90_per_10k) metric to find undervalued players
• Assigned players one of six Region Groups based on their nationality.

3

4

Methods

Modeling

• Filtered data by players who played more than 1,000 minutes and made less than $2 million
• Fit a linear model to predict ga_per_90_per_10k based on player ages, positions, and nationalities
• Used Generalized Additive Models to identify non-linear relationships between salary and G+ and age and G+ per 90 per $10,000

5

6

7

8

Discussion

9

Defense wins Championships (Maybe)

• While our team level models found that teams spending a higher proportion of their salary on defense did better, our G+ player level models found Strikers and Wingers to be slightly undervalued
• We lean towards defenders being undervalued given G+ bias towards attackers and that defenders are cheaper

​

Europeans, South Americans, Asian/Oceanic and Older Players are Overvalued (until they’re really old)

• Being European, South American, Asian/Oceanic or being older all have negative effects on Goals Added per 90 minutes per 10,000 dollars spent

So what should MLS clubs do?

Younger, Domestic Players are Undervalued. Find Them.

10

1. Academies

2. SuperDraft

Case Study: Patrick Agyemang

Drafted in 2023:

11

In 2024…

Salary: $71,401

Goals Scored: 10

12

In 2025…

​

13

14

Thank You!

​

Questions?

15

Appendix

16

17

18

19

Methods

​

20

Team Modeling

• Split up each teams roster into the percentage of their total salary they pay different slots of players
• Split up into 6 groups of 3 of the top 18 paid players
• Predicted xG difference using linear regression, regularization models, and XGBoost
• Trained the models on data from 2021-2023, tested it on 2024 and used bootstrapped cross-validation to evaluate model performances

21