Two different problems
Team games raise two separate questions. The first is how to predict a match: given the players on each side, who is more likely to win? The second is how to assign credit afterwards: given the result, how should each player’s rating change? Both usually start from the same step, which is combining individual ratings into a single team figure. This article is about that combining step and where it goes wrong.
The simplest method: the average
The most common approach is to take the mean of the players’ ratings. It is easy to explain, works for any team size, and gives sensible numbers for similar players. If five players are rated 1500, 1550, 1600, 1650 and 1700, the team figure is 1600.
The weakness is that an average hides how ratings are spread. Consider two five-player teams:
- Team A: 1600, 1600, 1600, 1600, 1600. Mean 1600.
- Team B: 2000, 1800, 1400, 1400, 1400. The sum is 8000, so the mean is also 1600.
The system sees these as equal. In practice they may not be, and which side is favoured depends on the game. In some games one standout can carry a lineup; in others the weakest player gets targeted and the team is only as good as its soft spot. A plain mean cannot express either idea.
Sums and the effect of team size
A sum behaves like a mean when both teams have the same number of players, since the comparison is unchanged. It matters when team sizes differ, such as a short-handed side after a disconnect or a mode with variable group sizes. Then the sum favours the bigger team, which may or may not be what you want, and the mean ignores the missing player completely. Neither treats a missing player as a real loss of capability, so systems that care about this usually add a specific adjustment.
Weighted blends
To reduce the problems of a flat average, some systems weight players unequally. Common options include:
- Weighting by role. In games where roles differ in impact, a role may carry more weight in the team figure.
- Weighting toward the top. Giving the strongest players more influence reflects the idea that a star changes outcomes.
- Weighting toward the bottom. Giving the weakest player more influence reflects the idea that a team gets exploited at its weak point.
Each choice is a claim about how the game works, and a wrong weighting is no better than a plain average. Picking weights is a modelling decision that should be tested against results, not assumed. A blend that fits one game can mislead in another.
Learning the combination from data
A more careful approach is to let results decide. Instead of choosing a formula in advance, the system fits a model that predicts outcomes from the players present, then checks whether its win probabilities match reality. Individual ratings are adjusted so that the team prediction improves over many games. This is the idea behind rating systems designed for teams, where each player has a rating and uncertainty and the team result is treated as evidence about the group.
This handles uneven contributions better, but it needs a lot of games per player, and it still assumes that a player’s strength is the same on every team.
Where the shortcuts fail
Synergy is invisible. Players who communicate well, or whose roles fit together, often do better than their separate numbers imply. Good team communication does not appear in any individual rating, and neither does a drafted composition that works well together. A sum of parts cannot capture a whole that is more than that.
Roles and matchups differ. A player rated highly in one role may be average in another, and a single number covers up the difference. Map pools and matchup structure can also change who has the advantage.
Stacks versus solo queue. A premade group tends to coordinate better than five strangers with identical average ratings. Systems that ignore this will consistently overrate the strangers or underrate the stack, so some ladders apply a separate adjustment for grouped players.
Credit assignment is crude. After a loss, the system lowers everyone, even if one player did most of the damage. Over many games, this tends to even out, because strong players win more often, but any single update is a rough guess. This is partly why individual ratings in team modes move more slowly and can feel unfair. Blame bias, the human tendency to pin a loss on teammates, mirrors the system’s own problem of not knowing who was responsible.
Rosters change. A team rating built from last month’s lineup says less when two players have been swapped. Esports teams that change a member are usually treated as a partly new entity, since the combined history only partly applies.
Small samples. A new team has few games together, so its team figure rests mostly on individual numbers. That is reasonable as a starting guess, but it is just a guess until results accumulate. Anyone tracking a team across scrims should treat early results with the same caution as any small sample.
A practical way to use team figures
For prediction, a team figure is a useful first approximation, especially when the teams are roughly balanced. The estimate is weaker when the lineups are lopsided, when a role is clearly mismatched, or when one side has played together far longer. In those cases, a combined number should be read as a starting point and checked against actual head-to-head records and recent form rather than treated as a final answer.
The honest summary is that every combination rule embeds an assumption about how individual ability turns into team strength. Averages assume equal influence, weighted blends assume a particular shape, and learned models assume enough data. Knowing which assumption is in play helps explain why the same lineup can be rated differently by different systems.