View on GitHub

Baron vs Dragons

A DSC80 project

Dragons or Baron? Modeling Objective Control and Win Probability in League of Legends

Author: Joshua Huang

Website: https://joshuaamm7.github.io/Project4-Baron_OR_Drags/

Introduction

Background

Competitive multiplayer games like League of Legends involve teams competing to control strategic objectives throughout a match. Among these objectives, dragons provide permanent team-wide buffs that can significantly influence the outcome of the game. Different dragon types grant different advantages, such as increased damage, movement speed, or durability. In addition to dragons, Baron Nashor provides a powerful temporary buff that can shift the momentum of a match in a single skirmish.

Because professional and ranked players often prioritize these objectives, understanding how different objectives influence the probability of winning can provide insight into effective gameplay strategies.

Research Question

Between Baron Nashor and Elemental Dragons, which objective has a stronger relationship with a team’s probability of winning in League of Legends?

Both objectives provide powerful advantages. Dragons grant permanent stacking buffs that accumulate over the game, while Baron grants a temporary but highly impactful buff enabling teams to push lanes and close out games. By comparing the relationship between these objectives and match outcomes, this project determines which is the stronger predictor of victory.

Dataset

The dataset comes from Oracle’s Elixir and contains years of professional League of Legends esports match data. The raw dataset has 120,636 rows and 165 columns. Each match contributes 12 rows: 10 player-level rows and 2 team-level rows. Since this project focuses on objective control and team outcomes, only the 2 team-level rows per game (20,106 rows total across 10,053 games) are used for analysis.

The key columns used in this analysis are:

Column Description
gameid Unique match identifier
result Match outcome: 1 = win, 0 = loss
dragons Number of dragons secured
barons Number of Baron Nashor buffs secured
elders Number of Elder Dragons secured
heralds Number of Rift Heralds secured
void_grubs Number of Void Grubs secured
atakhans Number of Atakhans secured
teamkills Total kills by the team
totalgold Total gold earned
damagetochampions Total damage dealt to enemy champions

Data Cleaning and Exploratory Data Analysis

Cleaning Steps

The raw dataset required several cleaning steps before analysis:

  1. Column selection: Only columns relevant to team-level objective control and match outcomes were kept.
  2. Filtering to team rows: Only rows where position == "team" were retained, reducing the dataset from 120,636 rows to 20,106 rows (2 rows per game).
  3. Handling missing values: The elders column had 1,634 missing values and atakhans had 1,686 missing values. These were filled with 0 where appropriate, under the assumption that a missing value indicates the objective was never secured.
  4. Feature engineering: Difference columns (e.g., dragons_diff, barons_diff, totalgold_diff) were computed by subtracting each team’s count from their opponent’s count within the same game. Boolean advantage columns (dragon_adv, baron_adv) were also created.

After cleaning, the dataset had 20,106 rows with no remaining missing values in any key objective column.

Sample of cleaned dataset (first 5 rows):

gameid teamname result dragons barons heralds totalgold
LOLTMNT03_179647 IziDream 0 0.0 0.0 0.0 42,255
LOLTMNT03_179647 Team Valiant 1 2.0 1.0 1.0 53,936
LOLTMNT06_96134 Esprit Shōnen 1 3.0 1.0 1.0 64,669
LOLTMNT06_96134 Skillcamp 0 2.0 0.0 0.0 50,679
LOLTMNT06_95160 Karmine Corp Blue Stars 0 0.0 0.0 0.0 51,389

Univariate Analysis

Distribution of Dragons Secured

Most teams secure between 1 and 3 dragons per game. Very few teams reach 5 or more, and securing 7 dragons is extremely rare.

Distribution of Barons Secured

Baron is a rarer objective, most teams secure either 0 or 1 per game. Multiple barons in a single game is uncommon and signals a long match.

Bivariate Analysis

Dragons Secured vs Match Outcome

Teams that win typically secure 2–4 dragons, while losing teams often secure only 1–2. However, the distributions overlap considerably, meaning dragons alone do not guarantee victory.

Barons Secured vs Match Outcome

The contrast here is much obvious. Most losing teams secure 0 barons, while winning teams almost always secure at least 1. This suggests Baron control is significantly more decisive than dragon control.

Aggregate Table: Win Rate by Objective Advantage

The table below shows the win rate for teams that secured more of each objective than their opponent, compared to teams that did not. The win-rate difference measures how strongly each objective predicts victory.

Objective Win Rate (More) Win Rate (Less/Same) Win Rate Difference
Barons 0.905 0.249 0.656
Dragons 0.830 0.254 0.577
Atakhans 0.796 0.262 0.534
Elders 0.870 0.487 0.383
Heralds 0.662 0.339 0.322
Void Grubs 0.600 0.419 0.181

Teams with Baron advantage win over 90% of the time. Baron has the largest win-rate gap of all objectives, followed by dragons.

Win Rate vs Dragons Secured

Dragons Secured Win Rate
0 3.6%
1 18.0%
2 46.8%
3 69.4%
4 90.2%
5 94.9%
6 97.1%
7 100.0%

Win rate increases sharply with each additional dragon, suggesting that while individual dragons are not decisive, accumulating them strongly correlates with winning.

Win Rate vs Barons Secured

Barons Secured Win Rate
0 21.9%
1 84.1%
2 83.4%
3 83.3%
4 50.0%

Securing even a single Baron pushes a team’s win rate above 84%. Teams that fail to secure any Baron win less than 22% of the time.

Assessment of Missingness

MNAR Analysis

One column with non-trivial missingness is atakhans, which has 1,686 missing values. The missingness likely occurs because this objective does not appear in every match. It may be absent in certain game patches or in games that end before the objective spawns. Because the probability that a value is missing may depend on factors not directly observed in the dataset (such as game patch or match duration), the missingness of atakhans is plausibly Missing Not At Random (MNAR).

Permutation Test 1: Atakhans Missingness Does NOT Depend on Dragons

Null Hypothesis (H₀): The missingness of atakhans does not depend on the number of dragons secured. The distribution of dragons is the same for rows where atakhans is missing and where it is not.

Alternative Hypothesis (H₁): The missingness of atakhans depends on the number of dragons secured.

Test Statistic: Difference in mean dragons between missing and non-missing atakhans groups.

Result: Observed difference = −0.039, p-value = 0.266.

Since the p-value (0.266) is well above the 0.05 significance level, we fail to reject the null hypothesis. There is no strong evidence that the missingness of atakhans depends on dragons secured.

Permutation Test 2: Atakhans Missingness DOES Depend on Teamkills

Null Hypothesis (H₀): The missingness of atakhans is independent of the number of team kills.

Alternative Hypothesis (H₁): The missingness of atakhans depends on the number of team kills.

Test Statistic: Difference in mean team kills between missing and non-missing atakhans groups.

Result: Observed difference = −1.454, p-value ≈ 0.0.

Teams where atakhans is missing have approximately 1.45 fewer kills on average than teams where it is not missing. Since the p-value is essentially 0, we reject the null hypothesis. The missingness of atakhans depends on team kills. That tells us games with more kills are more likely to have non-missing atakhans values, which aligns with the idea that higher-kill, more prolonged games are more likely to include this objective.

Hypothesis Testing

Question

Is Baron advantage more predictive of winning than Dragon advantage?

Hypotheses

Null Hypothesis (H₀): Baron difference and Dragon difference are equally predictive of winning. Any observed difference in their win-rate gaps is due to random variation.

Alternative Hypothesis (H₁): Baron advantage is more predictive of winning than Dragon advantage.

Test Statistic

The test statistic is the difference between the win-rate gaps:

T = (Win Rate when Baron Advantage) − (Win Rate when No Baron Advantage) − [(Win Rate when Dragon Advantage) − (Win Rate when No Dragon Advantage)]

Results

The observed test statistic was 0.079, meaning the win-rate gap associated with Baron advantage is approximately 7.9 percentage points larger than the win-rate gap for Dragon advantage.

A permutation test with 2,000 simulations shuffled match results to generate a null distribution. The resulting p-value was approximately 0.0, meaning none of the 2,000 simulated statistics were as large as the observed value.

Since the p-value is far below the 0.05 significance level, we reject the null hypothesis. This provides strong evidence that Baron advantage is more predictive of match outcomes than Dragon advantage.

This result aligns with League of Legends gameplay mechanics: Baron Nashor provides an immediate, powerful team-wide buff that dramatically increases map pressure and siege capability. While dragons offer stacking long-term advantages, securing Baron often leads to immediate strategic advantages that can decide a match.

Framing a Prediction Problem

Prediction Problem

Can we predict whether a team will win a match based on their objective control and in-game statistics?

Response Variable

The response variable is result (1 = win, 0 = loss). This is a binary classification problem.

Features Available for Prediction

The model uses features measurable during gameplay: dragons, barons, heralds, elders, void_grubs, teamkills, teamdeaths, totalgold, damagetochampions, and monsterkills. Engineered features such as dragons_diff, barons_diff, totalgold_diff, and teamkills_diff are also included in the final model.

Evaluation Metric

Models are evaluated using accuracy, which is the proportion of correctly predicted match outcomes.

Baseline Model

Features Used

The baseline model uses two quantitative features: dragons and barons, representing the raw counts of each objective secured.

Model

I choose to use Logistic Regression as the baseline classifier because of its simplicity and easy interpretability for binary classification.

Performance

Baseline Accuracy: 0.8224

The model correctly predicted match outcomes approximately 82% of the time using only two features. This indicates that even raw objective counts carry strong predictive signal, confirming that neutral objectives play an important role in determining match results.

Final Model

Feature Engineering

To improve on the baseline, I engineered four additional features representing each team’s advantage relative to their opponent:

These relative features capture game dominance more directly than raw counts alone. A team with 3 dragons against an opponent with 0 is in a fundamentally different position than a team with 3 dragons against an opponent with 3.

Hyperparameter Tuning

The regularization strength C in Logistic Regression was tuned using GridSearchCV with 5-fold cross-validation over the values [0.01, 0.1, 1, 10, 100]. The best performing value was C = 100, which applies minimal regularization and allows the model to fit the training data more closely.

Performance

Final Model Accuracy: 0.9887

This is a substantial improvement over the baseline accuracy of 0.8224, which is an improvement of approximately 16.6 percentage points. The engineered difference features provide significantly more predictive information than raw objective counts, because they directly measure which team dominated objectives and resources during the game.

Fairness Analysis

Question

Does the model perform worse on short games (≤ 20 minutes) compared to long games (> 20 minutes)?

This is worth examining because short games may have fewer objectives secured and smaller gold differences, providing weaker predictive signals for the model.

Hypotheses

Null Hypothesis (H₀): The model performs equally well on long games and short games. Any observed difference in accuracy is due to random chance.

Alternative Hypothesis (H₁): The model performs worse on short games than long games.

Test Statistic

Difference in accuracy: (accuracy on long games) − (accuracy on short games)

Results

The observed accuracy difference was 0.261, meaning the model is approximately 26.1 percentage points more accurate on long games than on short games.

A permutation test with 2,000 simulations shuffled the long_game labels to generate a null distribution. The resulting p-value was 0.0165.

Since the p-value (0.0165) is below the 0.05 significance level, we reject the null hypothesis. The model performs significantly better on long games than on short games.

This result is expected: features like objective counts and gold differences become more informative as a game progresses. In longer games, the winning team typically builds larger and more distinguishable advantages in these statistics. In shorter games, fewer objectives are secured and differences are smaller, making predictions less reliable.