league_of_legend_matches_analysis

When behind at 15:00, which position is most forgiving?

Using all professional League of Legends matches in 2022 (150,588 player rows, 164 features), we study which role hurts team win chance the least when trailing at 15 minutes.

📅 Season: 2022 📊 150,588 rows 🧭 Snapshot: 15:00 🎯 Final metric: ROC AUC

Top finding (interactive) Deficit × Role (interactive) Confusion matrix GitHub repo

---

TL;DR

Support (SUP) sustains the highest comeback rate when a team is behind at 15:00 and the lane is not ahead in kills, consistently across mild → moderate deficits. Jungle (JNG) is typically second; Bottom (BOT) is most fragile. In severe deficits, Top shows an exception with relatively stronger resilience.

Our final model is a simple logistic regression using 15-minute gold and CS differences (ROC AUC ≈ 0.73 test). It outperformed more complex baselines here and remained interpretable.

Key Figures (interactive)

Comeback rate by role

SUP ≈ highest; JNG ≈ second; BOT ≈ lowest when behind at 15:00.

Deficit severity × role

Comeback rate rises as deficits get lighter; Top stands out only in severe cases.

Final model performance

Baseline logistic regression (gold/CS at 15:00) — ROC AUC ≈ 0.73 test.

More visuals (open)

• Counts by Position
• ECDF of 15′ Gold Difference (teams already behind)
• 15′ Gold Deficit by Position (box)
• Missingness test — by league (TVD permutation)
• Missing rate by league
• Missing rate by side
• Missingness test — by side (TVD permutation)
• Fairness: AUC Blue vs Red (permutation)

Data & Cleaning (brief)

• Scope: All professional matches in 2022. One row per player and two rows for team stats.
• Size: 150,588 rows, 164 columns. Focused columns: gameid, league, side, position, result, golddiffat15, deficit15, comeback15, has_more_kills.
• Filters: Kept matches with gamelength ≥ 900s; focused on player-level rows for position analysis; dropped leagues where 15′ timeline is systematically unavailable (MAR).

Hypothesis Test (role resilience)

H₀: Within each 15′ deficit bin, SUP and non-SUP have the same comeback rate. H₁: SUP has a higher comeback rate.

Method: Stratified (by deficit bins) permutation test on win rate differences; weighted by bin sizes.

Result: p-value ≈ 0.001 (B=1000) → reject H₀. Supports “SUP is most forgiving when behind.”

Prediction Task (15:00 → final result)

• Unit: Team-game; prediction made at 15:00.
• Target: result (win=1).
• Features: golddiffat15, csdiffat15.
• Model: Logistic regression (baseline & final).
  Why it won: Best ROC AUC on test and remains interpretable.

Fairness Check (Blue vs Red)

Metric: ROC AUC by side; Test: label permutation on group tags.

Outcome: AUC(Blue)=0.626, AUC(Red)=0.628, Δ=−0.0014; two-sided p=0.933 → no evidence of side bias.

See the permutation distribution ⟶

---