Today I’m launching a pair of new statistics that measure teams’ performances in the “early game” and the “mid/late game”. These stats use complex modeling to assign an “early-game rating” (EGR) and a “mid/late rating” (MLR) to each team, which lets us quickly compare teams’ performances using single, straightforward numbers. The higher a team’s EGR, the better they have performed in the early game. The higher their MLR, the better they have done in the mid and late game. It’s that simple!
These stats will start showing up on the OraclesElixir.com team stats tables soon.
Keep reading to learn about how EGR and MLR are produced and what they represent, or skip halfway down for lists of team ratings along with some discussion and interpretation.
Who is this for?
These stats will be most useful to moderate and casual fans, or to fans who don’t watch a particular region. Experts, analysts, and enthusiasts will still find EGR and MLR useful as starting points for team comparisons, and as a way to test and measure opinions, but if you watch most or all of the games in a given league, these stats might not tell you anything surprising.
What do EGR and MLR look like?
EGR is centered at 50, with scores higher than 50 being stronger and lower than 50 being weaker. Generally, an EGR of 60 or higher is very good, while a rating of 40 or lower is very poor.
MLR is centered at zero, and is expressed as a positive or negative number. Generally, an MLR +15 or higher is very good, while a -15 or lower is very poor.
EGR and MLR can be presented as individual numbers, or they can be presented together, e.g. as 60.0+10.0, where the first number is the EGR and the sign and second number represent the MLR. (In this form, the numbers are a deconstructed win rate. Check the Methods section below for more details.)
As an example, in the 2015 Summer split regular season, Team SoloMid had a 50.0 EGR and +11.1 MLR. This tells us that their early-game performances were only average, but their mid/late-game performances were fairly strong. SK Gaming had a 53.3-20.0 EGR/MLR. This tells us that their early game was a little better than average, but their mid/late-game performances were exceptionally weak.
The heart of these statistics is a logit, or logistic regression, model. (If you’re interested in seeing the actual coefficients and other output from the model, get in touch with me. I won’t bog us down with them here.)
In simple terms, the model uses data from almost 800 professional games to predict a team’s probability of winning the game based on the status of the game at the 15 minute mark.
The independent variables used in the model are gold difference at 15 minutes and Dragon difference at 15 minutes, with a dummy variable for map side (Red = 1) included as a control. The games included in this model come from the 2015 Summer split regular season, playoffs, regional tournaments, and promotion series, covering the NA LCS and CS, EU LCS and CS, LCK, LMS, and international wildcard events.
To help grasp how the model works, here’s how gold leads map to win rates at the 15 minute mark, depending on whether a team is ahead, even, or behind in Dragons.
For the stats buffs, I tested the model’s accuracy by performing 10-fold cross validation (thank you to @shdwfeather for her support and suggestions!). The model is approximately 76% accurate, a reasonable score.
Because of ongoing changes to the game, the model may be re-estimated each split. Further testing will show whether this is necessary. If re-estimation is necessary, it will lead to small mid-season and end-of-season adjustments to teams’ ratings.
For the sake of interest, I estimated separate models for the NA LCS, EU LCS, LCK, and LMS. The win probabilities for each region can be compared in the three charts below, which compare the regions when ahead by one Dragon, even in Dragons, or behind by one Dragons. Note: These regional models may be slightly less accurate due to smaller sample sizes.
In general, LCK teams were more effective than other regions at converting early leads into wins, producing higher probabilities with smaller gold leads. The NA LCS produced the highest probability of comebacks.
Dragons were relatively most important in the LMS, having a larger effect on win probability. But in the NA LCS model, a Dragon advantage at 15 minutes did not have a statistically significant effect on probability of winning, a pretty bleak indictment of North American teams’ inability to win games through objective control.
Producing EGR and MLR
In simple terms, EGR is a team’s average probability of winning their games as of the 15 minute mark. MLR is the difference between the team’s actual win rate and their “expected win rate” based on their EGR. Let’s take our TSM example: their 50.0 EGR means that, on average, they gave themselves a 50.0% chance to win games, based on how they played the first 15 minutes. Their +11.1 MLR means that their 61.1% win rate was 11.1 percentage points higher than the 50.0% they “should” have won.
Why 15 minutes?
I chose 15 minutes as the cutoff time for EGR for several reasons.
Using a multiple of 5 makes interpretations as straightforward as possible. This made 15 and 20 minutes the most attractive options, with 10 minutes as an outside possibility.
The 15 minute mark represents the “early game” very well. On average, in a game with standard lane setups, the first Tower falls around 12 or 13 minutes. That means the laning phase is still happening at 10 minutes, but is usually coming to an end by 15 minutes as the map begins to open up. By 20 minutes, regardless of lane setup, an average of 1.9 Towers have fallen for each team, so the mid game is typically well underway. The facts that a team can surrender as of 20 minutes, and that Baron Nashor spawns at 20 minutes, also suggest that 20 minutes is firmly in “mid-game” territory.
The range of EGRs at 15 minutes is narrower than the range of EGRs at 20 minutes. At 15 minutes, 19% of games are 90% or more in favor of the leading team. But at 20 minutes, that increases to 31% of games. In other words, basing EGR on 20 minutes means that comebacks are much less likely, since almost a third of the games are virtually already decided.
Results and Discussion
Note: Tables are sorted by regular season win rate.
Here are a few interesting takeaways from these ratings, as examples of how the numbers help tell the story of the teams’ performances.
In the NA LCS, compare Team Impulse’s EGR between the regular season and playoffs/regionals. When XiaoWeiXiao was suspended, the team was forced to rely much more heavily on Rush, and that led to them going all-in on the early game. Their EGR skyrocketed as a result, but their MLR plummeted as they failed to convert on the leads they were building.
Cloud9’s EGR during the regular season was respectable, but their MLR was second-worst in the league, ahead of only Team Dragon Knights. The team clearly missed Hai’s late-game shotcalling for the first half of the split, and needed some time to adjust to his return in the second half. But the real story comes from dividing the full Summer split (regular season and regionals) into Meteos’s games and Hai’s games: with Meteos, C9 scored a 44.5 -14.5 EGR+MLR, but with Hai they improved significantly to 51.7 +4.9.
Fnatic’s undefeated regular season produced monumental team ratings. Their early game was often criticized as their weak point, but Fnatic still had the highest EGR in the region, possibly because they beat up on weaker teams more effectively than others did. In the playoffs, where Fnatic faced the Unicorns of Love and Origen, they achieved an EGR of 55.1, still respectable but with a hint of weakness when facing tougher opposition.
Unicorns of Love
After Fnatic, the Unicorns of Love had the highest regular season MLR in North America or Europe, but their EGR was so atrocious that their ability to make comebacks was barely enough to squeak them into the playoffs. Once they hit the playoffs and regionals, where the competition stiffened, their pattern of play caught up to them: UoL’s EGR remained poor, and they couldn’t find the comebacks they had previously relied on.
The EGR and MLR statistics could be really valuable tools, and I’m looking forward to using them in my own analysis and hopefully seeing others pick up on them, too.
I’m open to questions and feedback on the methods, applications, and presentation of these numbers, so get in touch with me through a comment on this page or on Twitter (@TimSevenhuysen).
Thank you again to @shdwfeather for her statistical expertise and input, to @kierisi for producing the model visualizations, and to Joshua “Jatt” Leesman for his invaluable feedback and discussion as I developed the model and its interpretations.