Matchmaking generates more player frustration than almost any other system, and the underlying mathematics is comprehensible and explains most of it.
The rating model
Most systems descend from rating models developed for chess.
Each player has a number representing estimated skill. The expected outcome of a match is calculated from the difference between ratings.
Ratings adjust after each match based on the difference between expected and actual outcome — beating a stronger player gains more than beating a weaker one.
Which is elegant and works well for one-against-one games with clear outcomes.
The extensions
Modern systems add uncertainty tracking.
A new player's skill is unknown, so the system holds a wide uncertainty range and adjusts rapidly.
An established player's rating is more confident, so it adjusts slowly.
Which is why placement matches move rating dramatically and later matches move it slightly, and why returning after a long absence sometimes widens uncertainty again.
Team games
Where the model strains.
Attributing a team outcome to individual skill is genuinely difficult, since one player's contribution is confounded with four others'.
Systems handle this variously — averaging team ratings, weighting by role, incorporating individual performance metrics.
Individual performance metrics create their own problem, since players optimise for whatever is measured rather than for winning.
Which is why systems that reward statistics produce behaviour that damages team outcomes, and it is a recurring design failure.
The queue time trade-off
The constraint underlying most complaints.
Tighter skill matching produces better matches and longer queues, since fewer candidates qualify.
Looser matching produces faster queues and worse matches.
Systems generally widen their acceptable range as queue time increases, which means a long queue produces a worse match, exactly when the player is already frustrated.
Population size determines how severe this is, which is why matchmaking quality degrades as a game ages and at unpopular hours and in smaller regions.
Engagement-optimised matchmaking
The accusation players most frequently make and the hardest to evaluate.
Patents exist describing matchmaking systems that consider factors beyond skill, including engagement and monetisation.
Whether any specific game uses such a system is generally not disclosed, and studios have denied it in several cases.
What is verifiable is that the patents exist and that the mechanisms described are technically straightforward.
Which is why the suspicion persists in the absence of transparency, and transparency would resolve it.
The perception problem
A system working correctly produces a win rate near fifty percent for everyone at equilibrium.
Which means most players lose roughly half their matches regardless of improvement, since the system moves them up as they improve.
That is the intended behaviour and it feels like being held back, because improvement does not produce a rising win rate.
Rating rising while win rate stays flat is what progress looks like, and it is not intuitive.
Smurfing
Experienced players on new accounts, which corrupts the rating estimates of everyone they encounter.
Detection through behavioural analysis has improved and remains imperfect.
Faster initial calibration reduces the damage window, and account restrictions requiring investment before ranked play raise the cost of creating new accounts.
Neither eliminates it, and the population most affected is new players, which is the group games can least afford to lose.
Role and position matching
In games with distinct roles, matching must fill each one, which constrains the candidate pool severely.
Which produces long queues for popular roles and short ones for unpopular ones, and incentive systems offering rewards for filling scarce roles are the standard response.
Autofill, assigning players to roles they did not select, is the alternative and is consistently unpopular because performance out of role is worse.
Rating players separately per role addresses the accuracy problem and worsens the queue problem, since each rating has less data.
Party matching
Groups of friends playing together create a structural difficulty.
A coordinated group outperforms individuals of equivalent rating because communication is worth a measurable amount.
Which means systems apply a penalty to group ratings, and calibrating that penalty is genuinely hard since the advantage varies by group size and by game.
Matching groups only against other groups solves it and requires enough grouped players to be queuing, which is frequently not the case outside peak hours.
Decay and inactivity
Ratings that decay with inactivity keep leaderboards current and punish players who stop, which is a design choice rather than a mathematical requirement.
Regional servers
Matching within a region keeps latency acceptable and reduces the available population.
Which is why cross-region matching appears at unpopular hours, at a cost in connection quality that players notice immediately.
Server location within a region matters as much as the region itself, and games that disclose it allow informed choice.
Reporting behaviour
Separate from skill matching, systems that track conduct and match repeat offenders together have been deployed with reasonable results.
Which improves the experience for everyone else without requiring bans, and it is a straightforward application of the same matching machinery.