This passed weekend marked the end of the regular season of the Spring Split of EU & NA LCS in 2018. With an 18 game season and best of 1’s all around, there was a lot of talk about potential tiebreakers. Teams that started strong and ended weak, teams that started weak and ended strong, and generally inconsistent teams all roughly had a similar win rate throughout the split. This lead to a lengthy tie breaker process for playoff seeding. Not only does this put a strain on production resources, but it can be bad for players who are expected to maintain peak performance for nine or more hours in a day. Surely, there has to be a better solution.
As it turns out, there is a system that is suited for this task. It’s the Elo rating system, named after Arpad Elo. Elo calculates the probability of a victory by ranking players and comparing the difference between their ranks. The victorious player than takes a weighted number of points away from the loser based on their probability to win the game. A higher ranked player will take fewer points from a lower ranked one, whereas a lower ranked player will take more points if they win. This system is used in many games and was even used in League of Legends prior to Season Three.
In order to test it out, I create a simple tool in Unity where I could input the matches and return each team’s Elo rating. As you can see below, the first game of 2018 resulted in a loss for TSM and a victory for Team Liquid. Since both teams had the same Elo prior to the game, there was a 50% chance either could win. With a K-factor of 32, this has Team Liquid taking 16 Elo from TSM. The K-factor is a multiplier on the weight of each individual game. It is responsible for how sensitive the system is, and is usually set to a larger number when sample size is low or ratings are low.
After day one, all teams having played exactly once, each team rests at either a 1216 or 1184 Elo rating.
Anyone interested in looking at the full data is welcome to do so here. The first link leads to some processed data and the second link leads to what was exported by the calculator tool.
With all of the data put together, this is what the NA LCS looks like after week 1, day 2.
As you can see, the weighted victories and losses are already starting to take effect. TSM, which had the unique start of losing their second game to a team that had also lost on day 1 had suffered a more significant defeat the two other 0-2 teams in the week. Both Optic and Golden Guardians had lost their first game, but also lost their second game to teams which had won their first game. Of course, each team’s rating is not complete until they’ve all played each other at least once.
This process continues through to the first day of the fifth week and now you can see that there is more diversity in Elo ratings. Nearly every team has had a unique experience reaching this point and each win or loss cares about the sum of the wins or losses of their opponents and themselves. Each team had played each other at least once. The two teams in joint first, Cloud 9 and Echo Fox, have a clear dominating lead on the league in both ELO and ranking.
As we move into the end of the season, the discrepancy between Elo and ranking start becoming more noticeable. Although TSM, TL, CG, and C9 have the same win ratio, the teams they faced has shifted the rankings to favor those with more consistent victories in the second half. In the case of Team Liquid and Clutch Gaming, the battle for third favors Team Liquid by only a difference of 0.229 Elo. Echo Fox, who had been so dominant in the early and mid season, now accurately reflects their recent losses in a meaningful way.
We now have definitive records for for the entire season and a flat ranking for every team. One could argue for a tiebreaker between Team Liquid and Clutch, but that would be only a single tiebreaker game as opposed to the five we had on Sunday, the 18th.
All in all, there are some serious discrepancies between a weighted Elo system and the final ranking system. While there might be some flaws with the Elo system such as it’s propensity to have minor adjustments based on game order within a closed bracket and it’s inability to fully self correct in such a small sample size, I think that there is some merit to investigating another rating system for League of Legends eSports. Perhaps a more advanced system such as the Glicko 2 system, derived from Elo. With Riot pushing for more stability through franchising, this might be a more accurate way to evaluate rosters in the long term, across multiple splits.
