1.8 KiB
A Statistical Evaluation of Magic: The Gathering Commander Deck Performance
Overview
This project explores the use of statistical and machine learning models to evaluate the performance of Magic: The Gathering (MTG) Commander decks. By applying classification methods such as Decision Trees, Linear Discriminant Analysis (LDA), and Random Forests, the study investigates whether deck performance can be reliably predicted based on card compositions.
Key Features
- Data Collection: Tournament results and decklists were sourced from EDHTop16 and processed using Python and Selenium to create a comprehensive dataset.
- Models Applied:
- Decision Trees: Offers interpretable decision rules.
- Linear Discriminant Analysis (LDA): Examines linear relationships between features.
- Random Forests: Addresses overfitting through ensemble learning.
- Results and Analysis: The models achieved modest accuracy, highlighting the challenges of predicting deck performance due to high variability and data limitations in the Commander format.
Results Summary
While Random Forests achieved slightly better performance compared to Decision Trees and LDA, all models exhibited limited predictive power, with accuracy metrics generally around 65% or lower. The variability in Commander decks and small sample sizes constrained model stability and generalizability.
Future Work
Given the limitations of Commander data, the study suggests exploring more standardized formats like Modern MTG decks, where larger and more consistent datasets may improve model reliability.
Acknowledgments
This research was conducted as part of an advanced undergraduate course on linear statistical models at California State University, San Bernardino, in Fall 2024.