league-of-legends-analysis

---

Project maintained by thamimhussain Hosted on GitHub Pages — Theme by mattgraham

Does an Early Lead Decide the Game?

A 2022 League of Legends Professional Match Analysis 📈🎮

Name: Thamim Hussain

Email: thamimh@umich.edu

---

1 · Introduction

“League of Legends” (LoL) is the world’s most watched esport, drawing 30M+ peak viewers to its 2023 World Final.
The dataset from Oracle’s Elixir records every professional game: gold, experience, kills, objectives – all time-stamped.

Central question
If a team holds an assist and experience lead at 10 minutes, how likely are they to win the entire match?

Why you should care:

• Strategic insight: Coaches decide whether to play aggressively for early leads or draft for late-game scaling.
• Broadcast context: Casters can instantly gauge win probabilities for viewers.

Item	Value
Raw rows loaded	≈ 150K player+team rows
Rows after filtering	21312
Season analysed	2022 (patch 12.01)
Relevant columns	goldat10, opp_goldat10, xpat10, opp_xpat10, golddiffat10, xpdiffat10, killsat10, opp_killsat10, assistsat10, opp_assistsat10, firstblood, result

Column descriptions

• goldat10 / opp_goldat10: Total team gold at 10:00 (ms).
• xpat10 / opp_xpat10: Cumulative experience points at 10:00.
• golddiffat10 / xpdiffat10: Our minus their resource counts (positive = advantage).
• kills/assists at10: Scoreboard differentials.
• firstblood: Whether this team secured first kill.
• result: True = win, False = loss.

---

2 · Data Cleaning & Exploratory Data Analysis (EDA)

2.1 Cleaning Steps 🔧

1. Team-row isolation: Kept rows where position == 'team' or participantid ∈ {100, 200}; removes per-player duplicates.
2. Type coercion: Converted early-game numeric columns to float; invalid strings → NaN.
3. Derived diffs: If missing, created golddiffat10 = goldat10 − opp_goldat10, same for XP.
4. Feature engineering: Added killsdiff10, assistsdiff10 for modelling; cast firstblood → bool.
5. Row drop: Removed teams with missing gold/XP diff (≈ **«>>** rows, < x% of total).

Head of the cleaned dataframe (5 rows)

gameid	side	result	golddiffat10	xpdiffat10	killsdiff10	assistsdiff10
ESPORTSTMNT01_2690210	Blue	False	1523	137	3	5
ESPORTSTMNT01_2690210	Red	True	-1523	-137	-3	-5
ESPORTSTMNT01_2690219	Blue	False	-1619	-1586	-2	-2
ESPORTSTMNT01_2690219	Red	True	1619	1586	2	2
ESPORTSTMNT01_2690227	Blue	True	-103	813	-1	-1

---

2.2 Univariate Analysis

Gold diff @ 10 min is nearly normal with μ ≈ 0 and σ ≈ 1.8k gold; extreme early snowballs (>5k) are rare.

---

2.3 Bivariate Analysis

The diagonal trend shows gold and XP leads move together. Orange (wins) dominate the upper-right quadrant, supporting our hypothesis.

---

2.4 Interesting Aggregates

Win-rate by binned gold lead:

gold_bin	Win Rate
(-1885, -942]	0.25
(0, 942]	0.60
(2828, 3771]	0.92
…	…

Crossing +1k gold flips the matchup above 60% win chance; beyond +5k, victories are almost guaranteed.

---

2.5 Imputation

We did not impute: missing early-game stats only occur in “remake” or prematurely terminated games and were dropped to avoid skewing results.

---

3 · Framing the Prediction Problem

• Task type: Binary classification (win vs loss).
• Response variable: result, as it is the official match outcome recorded by tournament servers.
• Feature timing: Only metrics known at 10 min (no mid- or late-game leaks).
• Evaluation metric: ROC AUC – robust to class imbalance and grades ranking quality of predicted probabilities.

---

4 · Baseline Model 🛫

• Algorithm: Logistic Regression (interpretable linear classifier).
• Features (total = 2)

Name	Type	Encoding	Rationale
assistsdiffat10	Quantitative	Standard-scaled	Measures supportive pressure and skirmish tempo.
xpdiffat10	Quantitative	Standard-scaled	Captures lane dominance and level advantages.

There are 0 ordinal and 0 nominal predictors, so no one-hot encoding was required.

• Model pipeline: StandardScaler ➜ LogisticRegression(max_iter=1000, C=1.0)
• Train / test split: 75/25 stratified by outcome.

Metric	Value
Accuracy	0.686
ROC AUC	0.750

Is the model “good”?
A naïve coin-flip would score 0.50 accuracy and 0.50 AUC. Hitting 0.75 AUC means the model ranks a random win higher than a random loss 3 times out of 4 – respectable for only two features, but still misclassifies ~31% of held-out games. The linear boundary also assumes additive effects; real games likely exhibit non-linear interactions, so we expect headroom for improvement.

---

5 · Final Model 🚀

Added feature	Type	Why it should help before seeing results
golddiffat10	Quantitative	Gold directly buys combat stats; even a small 200g lead can secure first dragon or Herald.
killsdiff10	Quantitative	A kill grants 300g + tempo (death timer, map pressure) not fully reflected by assists alone.
firstblood	Nominal (binary)	First Blood awards extra 400g and often snowballs lanes; encoded as 0/1.

Features are scaled through StandardScaler (numeric) while firstblood is already binary (no further encoding).

• Algorithm: Gradient Boosting Classifier (200 shallow trees, depth 2). Tree ensembles capture threshold effects (e.g., a 1k gold swing matters more than a 100g swing) and non-linear interactions.
• Hyper-parameter search: GridSearchCV (5-fold) over n_estimators {200,400}, learning_rate {0.03,0.05,0.08}, max_depth {2,3}. Best = 200 trees, lr 0.03, depth 2.

Model	Accuracy	ROC AUC
Baseline	0.686	0.750
Final	0.706	0.777

Why performance improved (data-generating perspective)
Gold is the universal currency for item power spikes; including golddiffat10 lets the model separate teams that traded kills for farm from those that did not. killsdiff10 captures raw elimination dominance, orthogonal to assists. firstblood encodes an early discrete snowball event worth ~400g that can’t be inferred solely from continuous differentials.

Tree ensembles exploit interactions like “XP lead and firstblood” being disproportionately strong, something the linear baseline cannot model. The resulting +0.027 AUC and +0.02 accuracy confirm these early-game signals sharpen win-probability estimates.

---