Snowball or Comeback?

Do early objectives decide professional League of Legends games? An analysis of 12,415 matches from 2022

Snowball or Comeback? How Early Objectives Shape Professional League of Legends

Live site: aryaharsh5559.github.io/lol-early-game-analysis/

Analysis of 2022 professional League of Legends matches (Oracle’s Elixir data).

Introduction

In professional League of Legends, the first 15 minutes of a match are often called the “laning phase” — a critical window where teams jockey for gold, experience, and map objectives. Analysts and players constantly debate whether early objectives like first dragon are truly game-changing or merely symptoms of a team that is already ahead. Our project investigates:

Do early-game objectives provide a genuine strategic advantage independent of a team’s overall gold position, and how reliably can match outcomes be predicted from the first 15 minutes of play?

We use the 2022 Oracle’s Elixir dataset, which contains 24,830 team-level rows (two per match) from 12,415 professional matches across 55 leagues worldwide. The dataset spans the entire 2022 competitive season (January–December) and includes detailed early-game state data alongside objective flags.

Relevant columns:

Column	Description
`result`	Match outcome (1 = win, 0 = loss)
`golddiffat15`	Gold lead/deficit at 15 minutes
`xpdiffat15`	XP lead/deficit at 15 minutes
`csdiffat15`	Creep score lead/deficit at 15 minutes
`firstblood`	Whether this team secured first blood (1/0)
`firstdragon`	Whether this team claimed the first dragon (1/0)
`firsttower`	Whether this team destroyed the first tower (1/0)
`gamelength`	Total game duration in seconds
`side`	Blue or Red side
`league`	Tournament/region (e.g. LCK, LPL, LEC)

Data Cleaning and Exploratory Data Analysis

Data Cleaning

We performed the following cleaning steps on the raw dataset:

Filtered to team-level rows: The raw dataset contains 148,980 rows (12 per game: 10 player rows + 2 team summary rows). We kept only the position == 'team' rows (24,830 rows), since our analysis focuses on team-level outcomes and aggregate early-game statistics.
Parsed dates: Converted the date column from strings to proper datetime objects.
Created readable game length: Converted gamelength from seconds to minutes (gamelength_min = gamelength / 60).
Assessed missingness in 15-minute columns: The columns golddiffat15, xpdiffat15, and csdiffat15 are missing for 15.2% of team rows — these correspond to games that ended before the 15-minute mark. We created a clean subset teams_15 (21,044 rows) with complete 15-minute data for modeling.
Engineered early objective count: Created early_obj_count as the sum of firstblood, firstdragon, and firsttower (range 0–3), capturing how many early objectives each team secured.

Head of the cleaned DataFrame:

gameid	league	side	result	gamelength_min	golddiffat15	firstblood	firstdragon	firsttower
ESPORTSTMNT01_2690210	LCKC	Blue	0	28.6	107	1	0	1
ESPORTSTMNT01_2690210	LCKC	Red	1	28.6	-107	0	1	0
ESPORTSTMNT01_2690219	LCKC	Blue	0	35.2	-1763	0	0	0
ESPORTSTMNT01_2690219	LCKC	Red	1	35.2	1763	1	1	1
8401-8401_game_1	LPL	Blue	1	22.8	NaN	0	0	1

Univariate Analysis

The distribution of professional game lengths is roughly normal with a right skew, centered around 30–32 minutes. Most games last between 25 and 40 minutes, with very few ending before 20 minutes or lasting beyond 45 minutes.

The gold differential at 15 minutes (golddiffat15) follows a roughly symmetric, bell-shaped distribution centered at zero, with most values falling between −3,000 and +3,000 gold. The symmetry is expected — every positive gold lead for one team is a matching deficit for the opponent. The tails extend to ±8,000+, representing rare blowout early games.

Bivariate Analysis

The box plot below shows that teams with a positive gold differential at 15 minutes are far more likely to win. The median gold differential for winning teams is roughly +1,500 gold, while losing teams sit at approximately −1,500 gold. However, there is significant overlap — many teams with early deficits still win, motivating our investigation into whether specific objectives matter beyond raw gold.

Win rate also varies by early objective control. The chart below shows that teams securing more early objectives (first blood, first dragon, first tower) win at dramatically higher rates — from 24.1% with zero objectives to 79.9% with all three. This compounding effect suggests that early objectives have a reinforcing strategic impact.

Interesting Aggregates

The table below shows first dragon’s win-rate impact in close games (gold differential at 15 min within ±1,000) versus decisive games (gold differential > 3,000). Crucially, first dragon still confers a 13.2 percentage-point advantage even in close games, suggesting an independent strategic value rather than simply being a byproduct of an existing gold lead.

Game Type	First Dragon	Win Rate (%)	Games
Close (\|GD15\| ≤ 1k)	Got First Dragon	56.6	2849
Close (\|GD15\| ≤ 1k)	No First Dragon	43.4	2849
Decisive (\|GD15\| > 3k)	Got First Dragon	60.3	3313
Decisive (\|GD15\| > 3k)	No First Dragon	39.7	3317

Assessment of Missingness

NMAR Analysis

We believe that golddiffat15 is not NMAR. Its missingness is almost entirely explained by gamelength — games that ended before the 15-minute mark simply have no 15-minute statistics to record. Since game duration is observed, this makes the missingness MAR (Missing At Random), conditional on game length.

A column that could plausibly be NMAR is url (the match replay link). Whether a replay URL is recorded might depend on factors intrinsic to the match itself — for instance, replays from games with technical issues or controversial rulings may be less likely to be archived. Since these factors relate to the content of the missing data rather than other observed columns, this would constitute NMAR. Obtaining additional data about tournament broadcast and archival policies could explain the missingness and reclassify it as MAR.

Missingness Dependency

We tested the missingness of golddiffat15 against two other columns using permutation tests (10,000 permutations, α = 0.05).

Test 1: Missingness depends on gamelength (p < 0.0001)

Games with missing golddiffat15 have a mean length of 30.91 minutes, compared to 31.74 minutes for games without missing values. The observed difference of −0.83 minutes is highly significant (p < 0.0001), confirming that shorter games drive the missingness — games ending before 15 minutes cannot produce 15-minute statistics.

Test 2: Missingness does NOT depend on side (p = 1.0)

Blue and Red sides have identical missing rates for golddiffat15 (15.25% each). The observed difference is 0.0, with a p-value of 1.0. This makes sense: when a game ends before 15 minutes, both teams (Blue and Red) have missing stats, so the missingness cannot depend on side.

Hypothesis Testing

We test whether first dragon provides a genuine strategic advantage, independent of overall gold state, using a permutation test.

Null Hypothesis (H₀): In close games (absolute gold differential at 15 min ≤ 1,000), teams that secured first dragon win at the same rate as teams that did not. Any observed difference is due to random chance.

Alternative Hypothesis (H₁): In close games, teams that secured first dragon win at a significantly higher rate, indicating that first dragon provides a genuine strategic advantage independent of gold state.

Test Statistic: Difference in win rates: WinRate(first dragon) − WinRate(no first dragon), among the 5,698 close-game team rows.

Significance Level: α = 0.05

Results: The observed difference is 13.16 percentage points (56.58% vs 43.42%). After 10,000 permutations of the firstdragon labels within close games, the p-value is < 0.0001. We reject the null hypothesis — there is strong evidence that first dragon provides a measurable competitive advantage even after controlling for close gold states at 15 minutes.

This is consistent with the game’s design: securing first dragon grants permanent stat bonuses (dragon soul stacks) that compound over time, providing a genuine strategic edge beyond the gold value of the objective itself.

Framing a Prediction Problem

We frame a binary classification problem: predict whether a team wins (result = 1) or loses (result = 0) using only information available at the 15-minute mark.

Response variable: result (1 = win, 0 = loss)

Features known at minute 15: golddiffat15, xpdiffat15, csdiffat15, firstblood, firstdragon, firsttower, and side. All of these are observable during a live broadcast at exactly 15 minutes into the game.

Evaluation metric: We use F1-score as our primary metric. F1 balances precision and recall, which is appropriate here because both types of errors matter: incorrectly predicting a win (false positive) and incorrectly predicting a loss (false negative) are equally costly for analysts trying to assess a team’s position. We also report accuracy, precision, and recall for completeness.

Baseline Model

Our baseline model is a Logistic Regression classifier implemented in a single sklearn Pipeline.

Features (7 total):

Quantitative (3): golddiffat15, xpdiffat15, csdiffat15 — continuous numeric values representing early-game resource differentials.
Nominal (4): side (Blue/Red, one-hot encoded), firstblood, firstdragon, firsttower — categorical features. side is encoded via OneHotEncoder(drop='if_binary'); the three binary flags are already 0/1.
Ordinal (0): None of our features have a natural ordering with more than two levels.

Preprocessing: Median imputation for numeric columns, most-frequent imputation for binary flags, one-hot encoding for side.

Train/test split: 80/20 with stratification on result (train: 16,835 rows, test: 4,209 rows).

Performance (test set):

Metric	Score
Accuracy	0.7505
F1	0.7517
Precision	0.7481
Recall	0.7552

The baseline model achieves ~75% accuracy, which is reasonable given that we only use 15-minute features. However, there is room for improvement by engineering features that capture how the early-game advantages interact (e.g., teams ahead in both gold and XP may be in a fundamentally different position than teams ahead in only one).

Final Model

Feature Engineering

We added three new features on top of the baseline’s seven, all computable at minute 15:

abs_golddiffat15 — the absolute value of the gold differential. This captures the magnitude of the gold gap regardless of direction. A team with a very large lead or deficit at 15 minutes is in a qualitatively different situation than a team in a close game, and this feature lets the model distinguish “lopsided” games from “close” games without knowing which side is ahead.
early_obj_count — the sum of firstblood + firstdragon + firsttower (0–3). As our EDA showed, win rate scales dramatically with the number of early objectives secured (24.1% with 0, up to 79.9% with all 3). This aggregate feature captures the compounding effect of objective control that individual binary flags cannot represent alone.
gold_xp_interaction — the product of golddiffat15 × xpdiffat15. Teams that lead in both gold and XP have reinforcing advantages (more items AND higher-level abilities), while teams ahead in one but behind in the other are in a more contested position. This interaction term captures that synergy.

Preprocessing

Numeric features (including the new engineered ones) are standardized with StandardScaler, which is important for logistic regression since it ensures regularization treats all features equally. Binary features and one-hot encoded side are imputed but not scaled.

Hyperparameter Tuning

We tuned the regularization strength C ∈ {0.01, 0.1, 1, 10, 100} with L2 penalty using 5-fold GridSearchCV scored by F1. We also evaluated Random Forest and Gradient Boosting classifiers, but the enhanced Logistic Regression outperformed both on cross-validation F1.

Best hyperparameters: C = 1, penalty = L2

Performance comparison (test set):

Metric	Baseline	Final Model	Change
Accuracy	0.7505	0.7515	+0.0010
F1	0.7517	0.7532	+0.0015
Precision	0.7481	0.7479	−0.0002
Recall	0.7552	0.7586	+0.0034

The final model improves F1 by 0.0015 and recall by 0.0034 over the baseline. While the improvement is modest, it reflects the inherent ceiling of 15-minute predictions — the gold differential alone captures most of the predictable signal, and the engineered features extract the remaining marginal information from objective interactions and gap magnitude.

Fairness Analysis

We investigate whether the final model performs equally well across different competition tiers.

Group X: Major leagues (LCK, LPL, LEC, LCS) — the four most prestigious professional leagues with the highest levels of competition (n = 410 test rows).

Group Y: Minor/regional leagues — all other leagues in the dataset (n = 3,799 test rows).

Evaluation metric: F1-score

Null Hypothesis: The model is fair — its F1-score for major-league games and minor-league games are roughly the same, and any differences are due to random chance.

Alternative Hypothesis: The model is unfair — its F1-score differs significantly between major and minor leagues.

Test statistic: F1(major) − F1(minor)

Significance level: α = 0.05

Results:

F1 (Major leagues): 0.7560
F1 (Minor leagues): 0.7529
Observed F1 difference: 0.0031
p-value (1,000 permutations): 0.911

We fail to reject the null hypothesis. The observed F1 difference of 0.003 between major and minor leagues is well within the range expected under random chance (p = 0.91). There is no evidence that the model performs unfairly across league tiers — it predicts match outcomes with roughly equal reliability regardless of the competition level.