League of Legends Live In-Game Prediction Model

BY Dan Manning

FILED UNDER ESPORTS

StatXP has recently extended our coverage of the League of Legends 2018 World Championships to include live prediction updates as matches are played out. We will be blogging and tweeting these updates during the semifinals and the world championship match. So now, you can see how team actions impact their odds of winning in real time!

The Model

Our live in-game prediction model is a logistic regression model that is based off a number of example models that forecast who will win sports competitions. Part of the work for this model was inspired by a University of Iowa master’s thesis called “A model for predicting the probability of a win in basketball”.

After learning about this model and others like it, I wondered - why can’t we model the probability of a win in League of Legends? And, why can’t I provide live updates as the model is updated during matches? The short answer is: I can.

Thus, the StatXP League of Legends Live In-Game Model was born.

For model inputs, the model the uses data from professional LoL leagues and international competitions. A special thanks to Tim “Magic” Sevenhuysen, Oracle’s Elixir, and Riot Games for making this data publicly available.

At first, given no information about the current status of the game, the model assumed that there was a 50% percent chance for each team to win. However, for the World Championship Matches, we have additional information about who might win based on each team’s past performance. In fact, StatXP has an entire Elo model dedicated to understanding each team’s past performance. Therefore I added each team’s Elo and their expected chance to win the match, prior to the start of the game, as a model input.

Next, our model systematically evaluated each game-related variable that I had match data for, including: team kills, team deaths, team assists, first blood time, first turret time, # of towers destroyed, bans, creep score at 10 minutes, first dragon time, first baron time, and many more variables. There were roughly 170 variables considered in total and the model selected the variables that were the most important predictors of match outcome. These selected variables are the ones the model tracks throughout the game and updates the live prediction each time one of the variables changes.

Finally, the outcome of the logistic regression model is a value between 0 and 1, which can be understood as the probability that a specific team is the winner of the match. The closer you are to 1 (or 100%), the more likely it is for your team to win!

 Cloud9 vs Afreeca Freecs Match 1 Outcome

Cloud9 vs Afreeca Freecs Match 1 Outcome

Model Application

Our team tested the model during match 5 of Invictus Gaming versus kt Rolster as well as during matches 1, 2 and 3 of Cloud9 versus Afreeca Freecs in the quarterfinals .

It’s been extremely exciting to watch the model edge upward or downward from the initial prediction based on a team’s performance. For example, In each quarterfinal game Cloud9 started down but was able to turn the tables on Afreeca Freecs. It was fun to watch that dynamic play out through the regression model and on Twitter with League of Legends fans.

We hope to see you for the semi-finals and the Lol Worlds 2018 Championship match. Stay on the lookout for our live model predictions using this model.