DSC80 — League of Legends 15-Minute Comeback Analysis

An end-to-end data science lifecycle project analyzing 150k+ professional LoL matches using hypothesis testing and logistic regression to predict comeback probabilities.

Course: DSC 80 – Practice of Data Science, UC San Diego  ·  Stack: Python, Pandas, Plotly, Scikit-learn  ·  Year: Summer 2025

GitHub → Live Demo →

Overview

An end-to-end data science lifecycle analysis on 150,588 professional League of Legends matches from the 2022 season. The core question: which player role — Top, Jungle, Mid, Bot, or Support — is most “forgiving” when a team is trailing in gold at the 15-minute mark? The project moves through rigorous data cleaning, statistical hypothesis testing, machine learning classification, and fairness evaluation to answer this with evidence.

Data Pipeline & Exploratory Analysis

  • Data Processing: Filtered 150,588 rows (164 features) for valid games (length ≥ 900s) and focused on player-level granularity to isolate role-specific metrics (golddiffat15, csdiffat15).
  • Missingness Mechanism Analysis: Discovered systematic missing data in 15-minute timeline statistics. Applied Total Variation Distance (TVD) permutation tests across professional leagues and map sides to classify missingness as Missing at Random (MAR), enabling principled handling before modeling.

Statistical Inference & Hypothesis Testing

  • Stratified Permutation Test: Formulated a null hypothesis that the Support (SUP) role and non-SUP roles share the same comeback rate. Stratified the test across deficit severity bins (mild to severe) and weighted by bin size to remove confounding from deficit depth.
  • Findings: Across 1,000 permutations the test yielded p ≈ 0.001, statistically confirming that the Support role is the most resilient position when behind at 15:00. The Bottom (BOT) role was the most fragile; Top lane showed unique resilience specifically in severe deficit scenarios.

Predictive Modeling & Fairness Evaluation

  • Baseline & Final Model: Trained and evaluated multiple classifiers, selecting a Logistic Regression model on 15-minute gold and CS differences. Achieved ROC AUC ≈ 0.73 on the test set — outperforming more complex baselines while retaining interpretability.
  • Fairness Check: Ran a label permutation test on Blue vs. Red side groups. AUC(Blue) = 0.626, AUC(Red) = 0.628, p-value = 0.933 — no evidence of side bias, confirming the model generalizes reliably across game environments.