Are League Games Decided at 20 Minutes, Even in Pro Play?
Introduction
In this analysis, I explore various aspects of League of Legends games at the professional level, particularly focusing on data collected at the 20-minute mark. League of Legends is (unfortunately) an incredibly popular game, with the highest level of competition drawing in tens of thousands of viewers around the world. With so many fans invested in outcomes of games, I wondered: “how feasible is it to predict the outcomes of games at the highest level?”
Data Cleaning and Exploratory Data Analysis
Before diving into the analysis, I cleaned the data to ensure that results could be more reliable. The key steps involved in cleaning the data are outlined below:
- Complete Data Selection:
- I included only games with complete data in the analysis. Complete data’s biggest advantage over partial data is that it includes lots of statistics and performance metrics for both teams AND individual players.
- Filtering for Games Above 20 Minutes:
- To avoid outliers and null values, I specifically only kept games where the game length exceeded 20 minutes. This was important because games shorter than 20 minutes will have missing data, and the amount of games at the highest level which end before 20 minutes are few and far between.
- Handling Missing Values:
- No imputation was needed because I filtered using the above critieria. Imputing data is particularily difficult for a dataset like this because the performance metrics are already dependent on so many factors, that trying to compute them would risk being very inaccurate.
Head of final cleaned dataframe:
| killsat20_blue_bot | killsat20_blue_jng | killsat20_blue_mid | killsat20_blue_sup | killsat20_blue_top | assistsat20_blue_bot | assistsat20_blue_jng | assistsat20_blue_mid | assistsat20_blue_sup | assistsat20_blue_top | deathsat20_blue_bot | deathsat20_blue_jng | deathsat20_blue_mid | deathsat20_blue_sup | deathsat20_blue_top | killsat20_red_bot | killsat20_red_jng | killsat20_red_mid | killsat20_red_sup | killsat20_red_top | assistsat20_red_bot | assistsat20_red_jng | assistsat20_red_mid | assistsat20_red_sup | assistsat20_red_top | deathsat20_red_bot | deathsat20_red_jng | deathsat20_red_mid | deathsat20_red_sup | deathsat20_red_top | golddiffat20_bot | golddiffat20_jng | golddiffat20_mid | golddiffat20_sup | golddiffat20_top | xpdiffat20_bot | xpdiffat20_jng | xpdiffat20_mid | xpdiffat20_sup | xpdiffat20_top | csdiffat20_bot | csdiffat20_jng | csdiffat20_mid | csdiffat20_sup | csdiffat20_top | gamelength |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 0 | 0 | 1 | 0 | 0 | 1 | 0 | 0 | 0 | 0 | 1 | 4 | 2 | 1 | 0 | 1 | 3 | 4 | 0 | 0 | 5 | 4 | 1 | 6 | 0 | 0 | 0 | 0 | 1 | 0 | -2460 | -943 | -647 | 1475 | 825 | -3714 | -942 | -290 | 3137 | 326 | -126 | 5 | 14 | 126 | 18 | 1760 |
| 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 1748 | -139 | -934 | 726 | -530 | 460 | 342 | -586 | 241 | 737 | 22 | -7 | -27 | -9 | -1 | 2523 |
| 1 | 0 | 0 | 0 | 1 | 0 | 1 | 1 | 1 | 0 | 4 | 1 | 5 | 3 | 1 | 3 | 3 | 5 | 0 | 3 | 4 | 8 | 4 | 12 | 2 | 1 | 0 | 0 | 1 | 0 | -723 | -59 | -2143 | -739 | -64 | -488 | 1653 | -1795 | -1370 | 507 | -9 | 67 | -14 | 2 | 8 | 2249 |
| 2 | 0 | 2 | 1 | 1 | 1 | 5 | 2 | 2 | 0 | 1 | 2 | 2 | 1 | 0 | 4 | 2 | 0 | 0 | 0 | 1 | 4 | 3 | 6 | 0 | 0 | 2 | 2 | 1 | 1 | -182 | -400 | 2302 | 87 | 1136 | 927 | 542 | 386 | -255 | 1038 | 18 | -5 | 20 | -2 | 36 | 1805 |
| 2 | 2 | 0 | 0 | 1 | 3 | 3 | 2 | 4 | 3 | 1 | 3 | 1 | 3 | 0 | 0 | 1 | 5 | 1 | 1 | 5 | 7 | 2 | 4 | 2 | 1 | 1 | 1 | 1 | 1 | -915 | -299 | 386 | 1874 | 1735 | -757 | -1120 | -574 | 2203 | 603 | -134 | -23 | 23 | 142 | 34 | 2075 |
Counterpick Winrates
One of the previous main focuses of this analysis was counterpick winrate. I calculated the winrate of champions when picked after the opposing champion of the same role. This has strategic implications because different champions often have weaknesses (think Fire Water Grass like Pokemon, but noticeably more complicated). This is useful for identifying champions that tend to perform well when played against certain opponents.
Here is an interactive plot displaying the counterpick winrates of champions:
In this plot, you can see that counterpicking can have a significant impact on chances of victory, depending on the champion used to counter.
To create this plot, I filtered for counterpicks, grouped by champions, then took the average of wins to get a win rate. Some rows of the aggregate table used for this are shown below:
| Champion | Winrate (%) |
|---|---|
| Naafiri | 85.71 |
| Ambessa | 83.33 |
| Kayn | 71.43 |
| Elise | 66.67 |
| Vel’Koz | 66.67 |
Pick Order by Role
Another important factor in team composition is the order in which champions are picked for each role, because it determines which roles on a team have to expose themselves first.
The following interactive plot illustrates the pick order across different roles:
The median values give lots of insight into how teams generally approach drafting roles. There’s a general trend towards picking the top laner later in the draft, while securing bot laner and jungler roles early. A basic interpretation of this could be that getting counterpicked in bot and jungle matchups is not as impactful as getting counterpicked in top, mid, or support.
Framing a Prediction Problem
The prediction problem this model aims to solve is: “Given performance metrics 20 minutes into a game, which team will win?”. This is a binary classification problem, since there are two teams and exactly one of them must win. The response variable is the outcome of the game. Accuracy is the metric used to evaluate the model because there should be a fairly even distribution between wins and losses across the dataset.
Baseline Model
For my baseline model, I used a Logistic Regression classifier to predict whether a team would win a game based on performance metrics taken at the 20-minute mark.
Features Used
The input features consisted of various champion-level metrics from both teams.
- Kills, Assists, and Deaths at 20 minutes per role
- Gold difference, XP difference, and CS difference at 20 minutes per role
- Game length
In total, there were 46 features. All of them were quantitative, so no encodings were needed.
Performance
On a training set of about 6000 games, the model had 78.72% accuracy. Its test accuracy was 80.25%. These results suggest the model is doing noticeably better than random guessing (which would probably be hovering around 50% +- 10%). The model is just under what I’d consider good, because there’s still plenty of factors to consider that could affect the outcome of a game.
Final Model
For my final model, I chose to incorporating feature selection, standardization, and hyperparameter tuning. Specifically, I added the following elements:
- StandardScaler: Normalize the feature values, ensuring the logistic regression model performs well since it is sensitive to the scale of the input data. This matters most for gamelength (which is in seconds) as well as the performance metrics for the support role, because it’s extremely rare for supports to have anywhere near as many resources as the rest of the roles on the team.
- SelectKBest (f_classif): Select the top 15 features most correlated with the target variable. This helps remove irrelevant or noisy features, improving generalization.
- LogisticRegression with GridSearchCV: Search for the best regularization strength (
Cparameter) across multiple values using 5-fold cross-validation.
I believe these features improved my model’s performance because the feature selection process explicitly chooses variables with the strongest statistical relationship to the outcome. This aligns well with the data generating process, where some champion performance metrics at 20 minutes are more predictive of game outcomes than others.
Performance
Using the same training and test sets, the final model had 78.55% training accuracy. Its test accuracy was 80.30%. The best performing hyperparamter value was C=0.1.
The final model showed an improvement over the baseline model in test performance, which indicates better generalization as the test set is unseen data to the model. I’d consider this an improvement because ultimately the goal of a model is to predict for unseen data.