Was it Over Before it Started?

A Statistical Analysis of Pre-Game Factors in League of Legends

Author: Viki Shi

General Introduction

Welcome to League of Legends—a video game developed and published by Riot Games, widely regarded as one of the greatest video games ever made. Beyond its popularity as a game, League of Legends is the world’s largest esport, with professional tournaments drawing hundreds of millions of viewers.

This analysis examines eight years of professional League of Legends matches to explore a central question:

Do pre-game decisions impact match outcomes?

In League of Legends, two teams of five players compete to destroy the enemy’s Nexus, their home base. With over 150 playable champions, in-game items, a gold economy, jungle monsters, and various team roles, the game is highly complex. However, for the purpose of this analysis, we focus solely on pre-game factors, which include:

By analyzing how these pre-game choices influence match outcomes, we aim to determine whether a game is significantly influenced before it even begins.

Roles and Champion Selection

There are five primary roles in League of Legends:

What is a Counter?

A counter is a champion that has a strategic advantage over another champion. Counters can:

What is a Ban?

Before the game begins, each player has the opportunity to ban one champion, preventing both teams from selecting that character. Bans are often used to:

What Are the Sides?

League of Legends is played on a symmetrical map called Summoner’s Rift, which has three lanes and a jungle area. Teams are randomly assigned to one of two sides:

Dataset Overview

This study analyzes 86,899 professional League of Legends matches, extracting key variables to measure the impact of pre-game choices.

Additionally, we scraped the following website https://www.counterstats.net/ to create our own dataframe consisting of:

Data Cleaning and Exploratory Data Analysis

Data Cleaning

Our dataset originates from Oracle’s Elixir, a comprehensive source of professional League of Legends esports data, covering thousands of matches from top-tier leagues such as the LCS, LEC, LCK, and LPL.

For this analysis, we selected data from the years 2017, 2019, 2020, 2021, 2022, 2024, and 2025. The excluded years only contained three bans per team, whereas we wanted to analyze games with all five bans consistently across all observations.

Preprocessing Steps

Dropped in-game statistics

Filtered out rows with equal PGA

Newly Added Features

To enhance our analysis, we created the following five additional columns:

Our added features enhance the analysis by quantifying the impact of pre-game decisions on match outcomes.

Our head of our final dataframe is as follows:

year gameid league teamname side ban1 ban2 ban3 ban4 ban5 pick1 pick2 pick3 pick4 pick5 num_counters_picked num_counters_banned PGA higher_PGA mean_champ_wr mean_team_wr result
2017 1506-1540 LPL I May Blue Syndra Malzahar Ashe Karma Poppy Maokai Kha’Zix Cassiopeia Varus Tahm Kench 0 2 2 1 0.508666 0.438776 1
2017 1506-1540 LPL Royal Never Give Up Red Camille Rengar Zyra Elise Rek’Sai Kled Lee Sin Ryze Caitlyn Nautilus 0 0 0 0 0.493086 0.598582 0
2017 1506-1541 LPL I May Blue Syndra Malzahar Ashe Rek’Sai Kha’Zix Maokai Lee Sin Corki Caitlyn Thresh 0 0 0 0 0.510833 0.438776 1
2017 1506-1541 LPL Royal Never Give Up Red Rengar Camille Varus Cassiopeia Orianna Trundle Rumble Ryze Jhin Zyra 1 0 1 1 0.492558 0.598582 0
2017 1507-1544 LPL Invictus Gaming Blue Jayce Elise Malzahar Kha’Zix Lee Sin Singed Rengar LeBlanc Varus Tahm Kench 0 1 1 0 0.49238 0.523126 1

Univariate Analysis

To begin our analysis, we examine the distribution of Pre-Game Advantage (PGA) across all games. A PGA of 0 indicates that a team neither counter-picked nor counter-banned, while a PGA of 10 would mean that a team countered every single champion possible.

Notably, PGA values do not exceed 8, due to the structured drafting process in professional League of Legends. The champion selection and banning process follows this sequence:

This format balances counter-picking opportunities, preventing one team from always picking first and the other from always countering. As a result, reaching a PGA of 10 is extremely rare, as it would require a team to deliberately choose champions that their opponents counter, which is unstrategic.

It seems that that there are more teams with PGA of at least 1, than none at all. Meaning, most teams either counter picked the enemy or banned one of their counters. This leads us to believe that players are considering the affect of counter matchups

Multivariate Analysis

Lets now see the affect on wins.

At first glance, PGA alone does not appear to strongly correlate with winning. However, this plot only shows general win rates for different PGA values—it does not assess whether having a higher PGA than the opposing team impacts victory rates.

To analyze this, we compare the win rates of teams with a higher PGA than their opponent.

The results reveal a 50/50 split, indicating that having a higher PGA does not significantly increase a team’s chances of winning.

Interesting Aggregates

To further explore the impact of pre-game factors on match outcomes, we calculated the mean aggregate values of key variables for both winning and losing teams. The results suggest that pre-game factors influence match outcomes.

result higher_PGA num_counters_picked num_counters_banned mean_champ_wr mean_team_wr blue red
0 0.49519 0.673699 0.643978 0.499192 0.471924 0.469407 0.530593
1 0.504811 0.675843 0.667331 0.500674 0.527778 0.530596 0.469404

While these aggregates support the idea that pre-game choices play a role in match outcomes, the differences are subtle. This suggests that while drafting strategy matters, in-game execution and adaptability likely have a greater influence on winning.

Assessment of Missingness

NMAR Analysis

Upon examining our dataset, we identified missing values in the following columns: teamname, ban1, ban2, ban3, ban4, ban5

The columns ban1 through ban5 contain missing values because players have the option to skip banning champions in a League of Legends game. Our preliminary analysis suggests that the missingness in these columns does not correlate with other observed variables, leading us to believe that the missing data is due to in-game choices rather than external factors, and as a result is Not Missing At Random (NMAR)

Missingness Dpendency

In this part, we are going to test if the missingness of teamname column depends on two other columns; year and league. The significance level for both permutation tests is 0.5, and the test statistic is Total Variance Distance (TVD).

First, we perform a permutation test on teamname and year

Null Hypothesis: Distribution of teamname when year is missing is the same as the distribution of teamname when year is not missing.

Alternative Hypothesis: Distribution of teamname when year is missing not the same as the distribution of teamname when year is not missing.

Below is the observed distribution of teamname when year is missing and not missing.

year team_name_missing = False team_name_missing = True
2017 0.0924056 0
2018 0.00296622 0
2019 0.134793 0
2020 0.160153 0
2021 0.197978 0.681818
2022 0.215255 0.318182
2023 0.00613622 0
2024 0.157504 0
2025 0.0328095 0

The p-value is less than the 0.5 significance level, and so we reject the null hypothesis. Thus, we can conclude that the missingness of teamname depends on the year.

Next, lets perform a permutation test on teamname and league

Null Hypothesis: Distribution of teamname when league is missing is the same as the distribution of teamname when league is not missing.

Alternative Hypothesis: Distribution of teamname when league is missing not the same as the distribution of teamname when league is not missing.

Below is the observed distribution of teamname when league is missing and not missing.

league team_name_missing = False team_name_missing = True
AC 0.000656643 0
AL 0.00400779 0
AOL 0.000634001 0
ASCI 0.001268 0
BIG 0.00600036 0
BL 0.0033285 0
BM 0.0039625 0
BRCC 0.00595508 0
CBLOL 0.0214654 0
CBLOLA 0.0122272 0
CDF 0.00246807 0
CISC 0.000588715 0
CK 0.0151707 0
CLS 0.00312472 0
CT 0.00267186 0
CU 0.00355493 0
DCup 0.00609093 0
DDH 0.0119781 0
DL 0.00310207 0
EBL 0.0124762 0
EBLPA 0.000656643 0
EGL 0.00151707 0
EL 0.00226429 0
EM 0.00697401 0
EPL 0.0026945 0
ESLOL 0.00896658 0
EU CS 0.00219636 0
EU LCS 0.00638529 0
EUM 0.0135178 0
EWC 0.000271715 0
GL 0.00323793 0
GLL 0.0153066 0
GLLPA 0.00144914 0
GPL 0.00113214 0
GSG 0.000815144 0
HC 0.00898922 0
HLL 0.00122272 0
HM 0.01268 0
HS 0.001268 0
HW 0.00303415 0
IC 0.00224164 0
IEM 0.000588715 0
KeSPA 0.00346436 0
LAS 0.0132687 0
LCK 0.0505615 0
LCKC 0.020922 0
LCL 0.00946472 0
LCO 0.00892129 0
LCP 0.00133593 0
LCS 0.0228693 0
LCSA 0.0232316 0
LDL 0.0830767 0
LEC 0.0233674 0
LFL 0.0184766 0
LFL2 0.00869486 0
LGL 0.000113214 0
LHE 0.0155557 0
LIT 0.00305679 0
LJL 0.0185672 0
LJLA 0.00135857 0
LLA 0.0159179 0
LLN 0.00430215 0
LMF 0.0167104 0
LMS 0.00978172 0
LPL 0.0715288 0
LPLOL 0.0159632 0
LRN 0.00163029 0
LRS 0.00165293 0
LSPL 0.0113667 0
LTA 0.000249072 0
LTA N 0.000430215 0
LTA S 0.000452858 0
LVP SL 0.0132687 0
MSC 0.000362286 0
MSI 0.00620415 0
NA CS 0.00201522 0
NA LCS 0.00889865 0
NACL 0.00830994 0
NASG 0.00194729 0
NERD 0.000815144 0
NEST 0.000498143 0
NEXO 0.0107554 0
NLC 0.0181822 0
NLC Aurora Open 0.00362286 0
OCS 0.0036455 0
OPL 0.0110271 0
OTBLX 0.00656643 0
PCS 0.0192464 0
PGC 0.00978172 0
PGN 0.011095 0
PRM 0.0181936 0.681818
PRMP 0.00426818 0.318182
RCL 0.00905715 0
RL 0.00124536 0
ROL 0.000543429 0
Riot 0.000317 0
SL (LATAM) 0.00321529 0
SLO 0.00724572 0
TAL 0.0135857 0
TCL 0.0230731 0
TPL 0.00156236 0
TSC 0.00178879 0
UGP 0.0020605 0
UKLC 0.00710986 0
UL 0.0158727 0
UPL 0.0176841 0
USP 0.00122272 0
VCS 0.0255185 0
VL 0.00317 0
WLDs 0.0125894 0

The p-value is less than the 0.5 significance level, and so we reject the null hypothesis. Thus, we can conclude that the missingness of teamname also depends on the league.

Hypothesis Testing

We aim to determine whether teams with a higher Pre-Game Advantage (PGA) have a significantly greater win rate than teams with lower PGA. This test allows us to assess whether pre-game factors, including champion picks and bans, impact a team’s likelihood of winning.

Null Hypothesis (H₀): The win rate of teams with a higher PGA is equal to 50%, meaning that PGA has no effect on winning. That is, having a higher PGA does not provide a statistical advantage in determining the outcome of the game.

Alternative Hypothesis (H₁): Teams with a higher PGA win significantly more than 50% of the time, indicating that pre-game decisions (picks and bans) contribute to winning.

Test Statistic : # games won by team with higher PGA / total number of games

Significance Level: We set α = 0.05 (5%), meaning we will reject the null hypothesis if the probability of obtaining the observed win rate (or more extreme results) under the null is less than 5%.

Here is a histogram containing the distribution of our test statistics under the null hypothesis:

Based on the hypothesis test performed, with a p-value of 0.0219, we reject the null hypothesis, as our significance level of 0.05. Meaning, under our trial, it appears that teams with a higher PGA win significantly more than 50% of the time, indicating that pre-game decisions (picks and bans) could contribute to winning.

Framing a Prediction Problem

We aim to predict the outcome of a League of Legends match using a binary classification model, where:

Our goal is to determine whether pre-game metrics and in-game factors influence match outcomes as the game progresses. To improve predictive performance, we explored various features and encodings.

We will use accuracy as the primary evaluation metric for our model, as a good accuracy will indicate that our model is making correct predictions at a high rate.

Baseline Model

To establish a starting point, we trained a baseline Decision Tree model on the following features:

Since our feature set included both quantitative and categorical data, we applied the following transformations:

OneHotEncoding: Used for the side feature since it is nominal and does not have an intrinsic order.

Standardization: Applied to our quantitative features (num_counters_picked and num_counters_banned) to bring them onto a similar scale, ensuring that no single feature dominates due to differences in magnitude.

These were all implemented into a single sklearn pipeline, and we then trained a Random Forest classifier with the following hyperparameters:

However, based on our exploratory data analysis and previous hypothesis testing, pre-game factors alone do not have a strong influence on match outcomes. As a result, our model only achieved 53% accuracy on the test set, which is only slightly better than a random guess (50%).

Our current model is not considered good due to its poor performance. A classification accuracy of 53% indicates that our model struggles to differentiate between wins and losses based on the provided pre-game features. This suggests that other in-game factors, such as player kills, deaths, assists, and economic decisions likely have a more significant impact on match outcomes.

Our baseline model only considered pre-game factors, which resulted in poor predictive performance. However, early to mid game events can significantly influence match outcomes. By incorporating gold, XP, creep score, and KDA metrics at the 10-minute mark, we redefined our classification problem to capture early-game momentum.

Our second model is a Random Forest Classifier trained on the following features:

Features from Data

Features Engineered

Note that we omitted the original features num_counters_picked and num_counters_banned, as - pga_diff rendered those columns slightly redundant.

In addition to one-hot-encoding side, to ensure fair weight distribution across features, we standardized all numerical values using StandardScaler(), which centers the mean at 0 and scales to unit variance.

These were all implemented into a single ski-kit learn pipeline, and we then trained a Random Forest classifier, finding the hyperparameters using GridSearchCV.

This model achieves an accuracy of 62%, marking a clear improvement over our baseline but still falling short of being highly reliable. Predicting a match outcome at the 10-minute mark proves challenging, which aligns with our understanding of early-game dynamics. During this phase, lanes remain largely isolated, limiting team interactions and strategic synergy. Additionally, major objectives—such as dragons, Rift Herald, and towers—are rarely contested before 15 minutes, meaning early advantages may not yet translate into definitive outcomes.

Final Model

To assess how predictive power evolves as the game progresses, we applied the same feature set but replaced 10 minutes with 25.

Features Engineered

The head of our dataset was the following:

side pga_diff kda_25 opp_kda_25 mean_champ_wr mean_team_wr golddiffat25 xpdiffat25 csdiffat25
Blue -1 1 0 0.513931 0.466667 1560 1638 52
Red 1 0 1 0.502757 0.555556 -1560 -1638 -52
Blue -1 3 3 0.497727 0.606061 776 943 39
Red 1 3 3 0.508189 0.397167 -776 -943 -39
Blue -1 2 1.33333 0.500533 0.444444 -59 690 -3

Our hyperparameters:

Our final model achieves an accuracy of 76%! This demonstrates that as the game progresses, we can predict the winning team with 76% accuracy using only kills, deaths, assists, experience, economy, and pre-game factors.

The accuracy is limited because our model makes predictions at the midway point of the game, without incorporating crucial late-game objectives such as turrets and major monsters (e.g., Baron Nashor or Dragon), which significantly impact the outcome.

Lets visualize how accuracy improves with gametime.

Fairness Analysis

The fairness analysis seeks to determine whether our model treats the two sides of the game (Blue and Red) fairly. Specifically, we aim to evaluate whether the model’s precision is roughly the same for both sides or if there is a statistically significant difference.

Alternative Hypothesis (H₁): Our model is unfair. Its precision for the Red Side is different from the precision for the Blue Side.

The results are as shown:

Our p-value is 0.2720, which means that under a significance level of 0.05, we fail to reject the null hypothesis. This indicates that our model is likely fair for both the blue and red sides, and any observed difference is likely due to random chance.

Conclusion

This study explored whether pre-game factors in League of Legends influence match outcomes, using statistical analysis and machine learning models. Our findings revealed that while drafting strategy plays a role, it does not significantly determine the winner. Instead, early- and mid-game performance metrics—such as gold difference, experience difference, and KDA prove to be stronger predictors of match success.

Our baseline model, relying solely on pre-game factors, achieved only 53% accuracy, highlighting the limited impact of drafting choices. However, incorporating early-game (10-minute) statistics improved accuracy to 62%, and our final mid-game (25-minute) model reached 76% accuracy, demonstrating that as the game progresses, prediction becomes more reliable.

Lastly, thank you for reading! I hope this study provides valuable insights into the complexities of League of Legends and the balance between strategy and execution.