By Amy Liu and Priya Gutta
View the Project on GitHub amyliiu/league-of-legends-analysis
This project uses the 2025 League of Legends (LoL) dataset from the Oracle’s Elixir dataset (https://oracleselixir.com/tools/downloads), which tracks all pro games played throughout the year. Our dataset was downloaded on March 25, 2025, and only includes matches played before that date. It contains detailed team and player-level data from games across multiple competitive tiers and regional splits. Each match contains statistics such as champion picks, total gold, kills, side selection (blue/red), and game results.
Specifically, we chose the 2025 LoL data because it reflects the most up-to-date game meta and team rosters. Because the game evolves quickly with frequent updates to champion and item abilities, using the most recent data ensures our analysis captures the current competitive landscape, making the results more applicable for understanding the true features that impact team wins in the modern meta.
In this project, our analysis will address the key question which features matter the most in determining team wins across a split? We will explore different elements that may affect the number of team wins, including:
Ultimately, the question we decided to explore further is: How do champion selection patterns, side choice, and total gold reflect team strategies and outcomes in professional League of Legends play?
Identifying features that influence team success provides insight into the strategic and performance-based factors that shape competitive e-sports. For example, a team’s variety in champion selection (champion diversity) is heavily dependent on their coaching staff and the champion compositions they explore. On the other hand, the average KDR reflects the mechanical skills of the players and their time spent in training. By analyzing these features, we uncover trends that drive a team’s success. This helps coaches, players, and administration make more informed decisions in scouting, preparation, and team development.
The dataset contains 24,373 rows of LoL match data, but we focus on the following columns that will help to answer our research question.
gameid: Unique identifier for each gamesplit: The competitive split (e.g., Spring, Summer)position: The position the player playedside: Team side (blue/red)teamname: Name of the teamteamid: Unique team identifierchampion: The player’s selected championresult: Game outcome (1 for win, 0 for loss)teamkills: Number of team killsteamdeaths: Number of team deathsdamagetochampions: Total damage dealt to championstotalgold: Total gold earnedWe performed the following data cleaning steps to ensure the dataset accurately reflects completed, competitive LoL games with meaningful information for analysis.
datacompleteness field was not marked as "complete". This ensures that only fully-played games’ statistics are included in our analysis."unknown team", likely due to missing metadata or scraping issues. We identified all gameids associated with these entries and removed all rows from those games to avoid partial or unbalanced comparisons.['gameid', 'split', 'position', 'side', 'teamname', 'teamid', 'champion', 'result', 'teamkills', 'teamdeaths', 'damagetochampions', 'totalgold'].team_df: rows where position == "team", representing overall team stats.player_df: rows for specific player game choices, such as champion selection.split ValuesNaN values were in the split column. Since the split indicates the tournament stage or league split the game belongs to (e.g., Spring or Summer), games without this information were excluded to avoid mixing in unstructured or potentially miscategorized matches. These rows were dropped using dropna(subset="split").Below is the head of the cleaned DataFrame:
| gameid | split | side | position | teamname | teamid | champion | result | teamkills | teamdeaths | damagetochampions | totalgold |
|---|---|---|---|---|---|---|---|---|---|---|---|
| LOLTMNT03_179647 | Winter | Blue | top | IZI Dream | oe:team:2799e04c7212d3c262467ef25427eda | Gnar | 0 | 3 | 13 | 20156 | 10668 |
| LOLTMNT03_179647 | Winter | Blue | jng | IZI Dream | oe:team:2799e04c7212d3c262467ef25427eda | Maokai | 0 | 3 | 13 | 4963 | 7429 |
| LOLTMNT03_179647 | Winter | Blue | mid | IZI Dream | oe:team:2799e04c7212d3c262467ef25427eda | Hwei | 0 | 3 | 13 | 13952 | 9032 |
| LOLTMNT03_179647 | Winter | Blue | bot | IZI Dream | oe:team:2799e04c7212d3c262467ef25427eda | Jinx | 0 | 3 | 13 | 6898 | 9407 |
| LOLTMNT03_179647 | Winter | Blue | sup | IZI Dream | oe:team:2799e04c7212d3c262467ef25427eda | Leona | 0 | 3 | 13 | 4174 | 5719 |
Because of our filtering, we had no rows that contained NaN values, and did not need to perform imputation. So, after performing data cleaning, we retained a high-quality dataset of 3120 rows and 1560 games for our analysis.
Early on, we explored questions such as: How do certain teams in Tier 1 pick their champions? Does a team’s side (blue or red) influence their selection? And does the split (e.g., Spring vs. Summer) affect how teams build their compositions? These questions guided our early data aggregations and visualizations, where we examined champion diversity, side preference, and changes in team strategies across splits.
We first analyzed the overall winning side frequency across all games:
From our figure, we see that the blue side won 140 more games than the red side. This is interesting, as in LoL games, the blue side always has the first pick in champion selection. We know that side assignment is random in games, so should coaches have players play more scrims on the red side?
Next, to understand how the team side affected champion selection, we analyzed the distribution of the top 10 most picked champions by side (blue or red).
Note that certain champions had a strong preference on one side. For example, Corki was picked 41 more times on the blue side than on the red, suggesting that it is valued as a first-pick option. This data could be used for champion bans, as players are aware of which champions their opponent is more likely to choose.
| Side | Loss | Win |
|---|---|---|
| Blue | 54,788.80 | 64,072.00 |
| Red | 53,651.00 | 64,909.50 |
This table shows us that winning teams earn significantly more total gold than losing teams, regardless of whether they play on the red or blue side. The gap highlights the importance of strategies such as farming, securing objectives, and controlling the map. Thus, we can assume resource control is a critical factor in determining game outcomes in LoL – so gold-related features might be useful in our predictive model.
| Side | Loss | Win |
|---|---|---|
| Blue | 11.0 | 20.9 |
| Red | 9.9 | 21.7 |
The “Average Team Kills by Side and Game Outcome” table shows a similar trend—winning teams secure up more kills,on average, than losing ones. This makes sense, as securing kills often leads to greater map control, gold income, and pressure. Note that this pattern holds true across both blue and red sides. So, kill-count-related features might also be useful in our model.
Thus, we aim to predict the number of wins a LoL team will achieve during a split using performance-based and strategic-based features. We are answering the question: Can we predict a team’s win rate during a split based on features such as average total gold, average kill/death count, champion diversity, side preference, and other average game performance stats? This is a regression problem because our target variable — the total number of wins per team per split — is a continuous number rather than a categorical label. We chose this target because win count is a direct indicator of a team performance across a season. The features used for prediction are:
All features are aggregated at the split level using data from previous splits. This way, they reflect team behavior and performance before the start of the split we’re predicting which simulates a real-world scenario. We want to forecast how well a team will perform in an upcoming split using only historical information and not provide any data from future games.
We use R² as our primary evaluation metric. This is because our primary goal is not just to predict win rate accurately, but to understand which features best explain a team’s performance over a split. R² directly measures how much of the variance in win rate is explained by the model, making it a more appropriate choice than MSE when interpretability and explanatory power are the focus. While MSE penalizes large errors and reflects raw prediction accuracy, it doesn’t tell us whether our model is learning meaningful patterns in the data. R², on the other hand, allows us to evaluate how well different feature sets and models capture the underlying structure of team success — which aligns more closely with our objective of identifying strategic factors that contribute to winning. By maximizing R², we prioritize models that offer clearer insights into what matters most for performance in competitive play.
We trained a baseline linear regression model to predict a team’s win rat in a given split.
We grouped the dataset by teamname and split, then computed average values per team per split to form these quantitative input features:
avrg_teamkills: Average number of kills per game.avrg_totalgold: Average total gold earned per game.The model includes 2 quantitative, 0 ordinal, and 0 nominal features. No encoding was necessary because all inputs were numeric. Note that the categorical variables of split and teamname were used only for grouping: they were not included as features in the model.
The target variable was the team’s win rate across each split, calculated as:
win_rate = number of wins / total number of games.
We split the data into a 70/30 training/test split and trained a linear regression model using scikit-learn. Here are the performance metrics on the test set:
| Metric | Value |
|---|---|
| MSE | 0.0264 |
| MAE | 0.1247 |
| R² | 0.401 |
We consider these results as a good starting point for modeling win rate considering we used just two stats. An R² of ~0.40 indicates that ~40% of the variance in win rate is explained by just two features, which is meaningful for such a high-variance outcome. Additionally, the low MAE (~12%) shows that predictions were generally close to the actual win rate, even with a simple model. The results also make intuitive sense: teamkills and totalgold are directly related to success in League of Legend.
To improve on the baseline model, we added the following features, chosen based on their relevance to team strategy in competitive LoL matches:
blue_ratio: The proportion of games played on the Blue side. The Blue side is traditionally seen as advantageous in pro play because it always gets first pick of champions. This ratio helps capture potential systemic advantages or preferences.champion_diversity: The number of unique champions a team played during a split. High diversity may indicate flexibility and strategic depth of a team, which could correlate with better performance.avrg_damagetochampions: The average damage dealt to champions. This statistic reflects heavily a team’s offensive capabilities.avg_kdr: The average kill-to-death ratio, calculated from teamkills and teamdeaths. A high KDR typically indicates team dominance during fights and good execution.These features reflect a team’s strategic style and adaptability, which go beyond raw metrics like kills and gold. Next, we trained four models to determine a one that would perform the best.
For our first model, we used the same linear regression model in the baseline model with a preprocessing pipeline:
avrg_teamkills and avrg_totalgold were scaled using StandardScaler to normalize their distribution.blue_ratio and avg_kdr were transformed using a QuantileTransformer to reduce skewness and account for outliers.champion_diversity and avrg_damagetochampions were left unscaled due to already reasonable distributions.| Metric | Baseline Model | Final Model |
|---|---|---|
| MSE | 0.0264 | 0.0131 |
| MAE | 0.1247 | 0.0795 |
| R² | 0.401 | 0.702 |
The R² improvement from 0.401 to 0.702 shows that our final model explains over 70% of the variance in win rate, compared to just 40% in the baseline. This is a significant improvement from our baseline model and shows that adding strategic features capture other aspects of the game that relate to its outcome.
We chose to explore LASSO next because it can effectively perform feature selection and improve generalization/reduce overfitting. We chose to tune the alpha hyperparameter for our Lasso model. Since Lasso applies L1 regularization, adjusting alpha helps control model complexity. A higher alpha increases the regularization strength, reducing overfitting and potentially zeroing out unimportant features. We used GridSearchCV with 5-fold cross-validation to search across a range of alpha values: [0.001, 0.01, 0.1, 1, 10]. The optimal value was found to be: alpha = 0.001
| Metric | Baseline Model | Final Model | Final Lasso |
|---|---|---|---|
| MSE | 0.0264 | 0.0131 | 0.0128 |
| MAE | 0.1247 | 0.0795 | 0.0789 |
| R² | 0.401 | 0.702 | 0.709 |
Compared to both the baseline and the earlier Linear Regression model, the Lasso model increased R² to 0.709, showing a stronger fit to the test data.
Additionally, we examined the importance of each feature based on its learned coefficient values. The Lasso model applies L1 regularization, which leads to some coefficients being zeroed out, effectively performing feature selection. Features with non-zero coefficients have a stronger impact on predicting win rate, while those with near-zero coefficients were considered less important.
The following table lists the features ordered by their absolute coefficient values:
| Feature | Coefficient |
|---|---|
avg_kdr |
0.148017 |
avrg_totalgold |
0.068793 |
avrg_teamkills |
0.032144 |
blue_ratio |
-0.019843 |
avrg_damagetochampions |
-0.000004 |
champion_diversity |
0.000000 |
From the table and plot, we can observe that avg_kdr has the largest positive coefficient, indicating it plays a significant role in the model’s prediction of win rate. Conversely, champion_diversity has a coefficient of zero, meaning it was dropped by the Lasso regularization process - this metric is unimportant.
For our final model, we selected a Random Forest Regressor due to its ability to model complex, non-linear relationships and handle interactions between features. We chose to tune n_estimators, max_depth, and min_samples_split, as these directly affect the complexity and performance of the model. We performed a grid search with 5-fold cross-validation to identify the best combination of these parameters, using R² as the scoring metric to maximize the proportion of variance explained by the model.
n_estimators: The number of trees in the forest. We tested values of 100 and 200 to compare performance between a smaller, faster ensemble and a larger one with lower variance.max_depth: The maximum depth of each tree. Limiting depth can help prevent overfitting. We tested depths of None (no limit), 5, and 10.min_samples_split: The minimum number of samples required to split an internal node. Increasing this value can reduce overfitting by preventing the model from learning overly specific patterns.n_estimators: 100max_depth: Nonemin_samples_split: 5| Metric | Value |
|---|---|
| MSE | 0.0097 |
| MAE | 0.0762 |
| R² | 0.781 |
The Random Forest model achieved an R² score of 0.781, indicating that it explains approximately 78% of the variance in the target variable (win rate), and performs better than any of our earlier models.
In Trial 2, we refined our hyperparameter tuning by focusing on a more detailed set of values for n_estimators, min_samples_split, max_features, and min_samples_leaf. From Trial 1, we determined max_depth should remain at None to avoid overfitting.
n_estimators: [100, 125, 150]min_samples_split: [5, 6, 7]max_features: ['sqrt', 'log2', None]min_samples_leaf: [1, 2, 3]n_estimators: 125min_samples_split: 5max_features: Nonemin_samples_leaf: 1| Metric | Trial 1 | Trial 2 |
|---|---|---|
| MSE | 0.0097 | 0.0098 |
| MAE | 0.0762 | 0.0764 |
| R² | 0.781 | 0.778 |
Despite a slightly better training cross-validation R² score (0.769 vs 0.767), the R² performance in Trial 2 (0.778) was marginally lower than Trial 1 (0.781). This suggests Trial 1 may generalize better, despite Trial 2 showing a stronger fit during training.
After evaluating multiple models, we selected the Random Forest model from Trial 1 as our final model. Although Trial 2 offered a similarly strong fit, the Trial 1 Random Forest model achieved the highest R² (0.781) among all models, indicating that it generalizes best while capturing the most relevant relationships in the data.
As a reminder, the prediction question we were trying to answer was: Can we predict a team’s win rate during a split based on features such as average total gold, average kill/death count, champion diversity, side preference, and other average game performance stats? We are trying to figure out which features contribute the most to success.
Using the Random Forest (Trial 1) model, we find that average KDR is the most important predictor of a team’s win rate. It has an importance of 0.8605, compared to the next highest feature importance, blue_ratio, with a importance of 0.0475.
Our work shows us that performance-based metrics such as KDR are more predictive of win rate than strategic elements like champion diversity or map side. This insight can guide analysts and teams in focusing on the most impactful aspects of play during a competitive split.