A multiplayer matchmaker can evaluate far more than a player’s rating. Match histories reveal how someone performs with different teammates, whether they leave games early, and how their participation changes after difficult sessions. Machine learning can use those records to estimate future behavior, including the likelihood that a player will stop returning to the game.
Those predictions create possibilities that traditional skill-based matchmaking was never designed to address. Developers could identify frustrating patterns, improve team composition, or test whether particular matchmaking decisions affect long-term participation. They could also allow retention targets to influence the opponents a player faces, with consequences for competitive fairness that may be difficult to detect from outside the system.
What MMR Measures, and What It Misses
Matchmaking Rating (MMR) remains a practical foundation for competitive play. Systems based on Elo, TrueSkill, and related statistical methods infer ability from previous results, often adjusting estimates as new evidence arrives. Depending on the system, a rating may also reflect individual performance, uncertainty, and the strength of the opposition.
The central task is predicting competitive ability well enough to assemble reasonable matchups. A rating should respond when someone consistently beats stronger opponents without swinging wildly after an ordinary losing streak. It also has to cope with limited histories, returning players, changing game modes, and parties whose combined performance differs from what their individual ratings suggest.
Microsoft’s TrueSkill 2 research, published in 2018, showed how additional information could improve these estimates. Its model incorporated factors including kills, experience, squad membership, quitting tendencies, and skill in other modes. In retrospective tests using Halo 5 matches, it predicted outcomes with 68% accuracy, compared with 52% for the original TrueSkill model.
TrueSkill 2 offers a useful distinction between richer skill estimation and behavior-based manipulation. Information about quitting or squad membership can make a rating more accurate without instructing a matchmaker to favor players who appear likely to leave. What developers choose to optimize after calculating a skill estimate is a separate question, and it determines whether behavioral information supports competition or begins to reshape it.
How Behavioral Prediction Works
Online games already record events needed to operate a multiplayer service, including match results, disconnects, party composition, and session activity. A predictive model can examine past records to identify patterns associated with a defined outcome. If the outcome is inactivity over the following two weeks, for example, the model learns which earlier signals tended to appear before that inactivity.
This process does not require a generative AI system. Gradient-boosted trees, statistical regression, and neural networks can all estimate probabilities from gameplay records. The choice of model depends on the available data, the speed required for decisions, and how readily developers can inspect mistakes.
Prediction also requires careful definitions. A player who closes the game after one match may be dissatisfied, short on free time, or simply finished for the evening. Someone disconnecting during a loss might be frustrated, but a network interruption produces much the same event record. Without context, a model can become very good at identifying a pattern while remaining uncertain about its cause.
Consider two competitors with identical MMR who have each lost four ranked matches. One regularly plays through long losing streaks and returns the next day, while the other has recently shortened their sessions and missed several usual play nights. A model might assign them different probabilities of returning, but that difference does not establish that either player needs an easier match. The prediction is an estimate of behavior, not a judgment about what constitutes fair opposition.
From Skill Matching to Churn Prediction
Churn generally describes players who stop participating within a defined observation period. A studio might track failure to return within seven or fourteen days, while another analysis might examine longer inactivity. Those definitions matter because abandoning an individual match and leaving a game for weeks are different behaviors, even if they sometimes occur together.
A prominent examination of retention-based matchmaking appeared in the 2017 paper EOMM: An Engagement Optimized Matchmaking Framework. The researchers proposed selecting pairings using estimates of how a match might affect subsequent disengagement, rather than treating skill parity as the sole objective. Their simulations, conducted with real player data from an Electronic Arts game, indicated that this approach could improve engagement compared with the alternatives they examined.
The paper is evidence that engagement optimization can be studied mathematically, but it is not evidence that EA deployed EOMM across its commercial games. It also does not establish that any named title deliberately hands players wins or losses to keep them playing. Those claims require evidence about an actual production matchmaker, not an inference drawn from a research proposal.
The idea nevertheless exposes a fundamental design choice. Two players may be equally suitable opponents on competitive grounds, yet a model could predict different retention outcomes for different pairings. If a matchmaking service begins ranking those pairings by expected engagement, it can reach a decision that differs from one based on skill, connection quality, and queue time.
Retention is a legitimate concern for a multiplayer operator. Falling participation can increase queues and make regional or mode-specific matchmaking harder to sustain. But an intervention that increases aggregate playtime still needs examination for who benefits, who absorbs the disadvantages, and whether the resulting matches retain their competitive meaning.
What Call of Duty’s Tests Actually Demonstrated
Activision’s 2024 technical report, The Role of Skill in Matchmaking, provides unusually concrete evidence about how matchmaking rules affect participation. It describes experiments that changed the weight of skill in Call of Duty: Modern Warfare III’s multiplayer matchmaking. The report also identifies connection quality and time to match as major considerations, showing that skill operates alongside other constraints.
In an early-2024 North American experiment, Activision loosened skill restrictions for a treatment group comprising half the tested player population. Skill remained part of the algorithm, while the comparison group received the standard configuration. After roughly a month, the company reported significantly lower 14-day return rates across nine of ten skill brackets for the treatment group.
The report also found significantly higher in-match quitting rates across eight of ten skill brackets. The highest-skilled group experienced better short-term results on some measures, while less-skilled players were more likely to encounter unfavorable conditions. Activision reported that separate tests with tighter skill restrictions negatively affected highly skilled players, illustrating the competing demands within a broad matchmaking population.
These were tests of skill constraints, not published demonstrations of AI selecting opponents according to individual churn predictions. The distinction matters because the results support a narrower, still meaningful conclusion: how a matchmaker groups players can influence both match completion and subsequent participation. They do not reveal every mechanism behind a commercial matchmaking system or settle how much behavioral optimization players should accept.
Team Behavior Adds Another Layer
Individual ratings struggle to capture everything that makes a team effective. Preferred roles, familiarity with a map, party coordination, and communication habits can influence outcomes even when the two teams have similar average skill. An advanced model could examine those patterns to estimate which combinations tend to work well together.
In a tactical shooter, a team of strong entry players may still struggle if no one consistently supports rotations or plays the objective. A matchmaker informed by role tendencies could potentially form better teams without relaxing its skill requirements. That would be a different application from arranging an easier opponent for someone judged likely to quit.
Behavioral systems also intersect with misconduct detection, particularly around harassment and disruptive communication. A 2025 study of toxicity in Modern Warfare III found effects on short-term engagement and the likelihood of similar behavior during a match. Those effects varied depending on whether toxicity came from teammates or opponents, whether teammates were in the same party, and whether the exposed player’s team won or lost.
More recent Activision-affiliated research published in July 2026 examined toxicity detection using chat alongside match context, including player and team information. It reported improvements over comparison methods in the research setting. These studies support the value of context-aware moderation, although neither establishes that such models determine who gets matched against whom.
When Engagement Goals Threaten Competitive Trust
A ranked player ordinarily expects the system to apply consistent criteria when selecting opponents. If the matchmaker secretly changes those criteria because someone is predicted to leave after another defeat, the resulting match becomes harder to interpret. A player might earn a win against deliberately weaker opposition without knowing that a retention objective influenced the pairing.
That scenario is hypothetical, not evidence that a particular game manipulates ranked results. Its importance lies in the incentive it exposes. A system rewarded for minimizing departures could favor decisions that keep vulnerable players active even when those decisions distribute competitive difficulty unevenly across the population.
Total playtime is also an imperfect proxy for a good experience. Someone may continue queuing because of a seasonal progression target while finding the matches unpleasant, or may finish a satisfying session after only two games. Developers who optimize a single engagement metric risk mistaking persistence for enjoyment, especially if their models overlook player surveys and match-quality measures.
There is a deeper statistical problem. A losing streak may precede inactivity without causing it; changes in available time, a new patch, or friends switching games might explain both. To determine whether a matchmaking adjustment actually helps, studios need controlled experiments and measurements that account for the players adversely affected by each decision.
Can Players Exploit Behavioral Models?
Competitive communities have long tested the boundaries of rating systems through reverse boosting, intentional deranking, and coordinated account manipulation. A matchmaker that responds visibly to distress signals could create a new incentive to manufacture them. Players might intentionally quit early or perform poorly if they believed those actions would produce weaker opponents later.
Even an incorrect belief could damage public matches. Players experimenting with supposed matchmaking tricks impose costs on teammates, and their behavior can distort the data used to train future models. Developers therefore need to distinguish ordinary inconsistency from deliberate manipulation without assuming that every unusual performance pattern indicates misconduct.
Models also face feedback effects after deployment. If an algorithm changes which opponents people face, their subsequent behavior may no longer follow the patterns observed in the original training data. A model that once predicted churn accurately could deteriorate as players respond to new incentives, leaving teams to re-evaluate both its predictions and its consequences.
Keeping behavioral risk estimates separate from the competitive rating calculation would make some of those problems easier to investigate. Administrators could examine suspicious departures or abrupt performance changes without silently converting them into opponent-selection advantages. The model’s findings would remain evidence for review rather than an automatic adjustment to competitive difficulty.
Transparency Without Publishing the Entire Algorithm
Matchmaking formulas are usually only partly visible to players, and developers have reasons to protect them from manipulation. Secrecy still comes at a cost. When a player experiences repeated lopsided matches, limited information creates room for theories about forced losing streaks, hidden punishment, or artificially favorable lobbies.
Developers can explain their goals without revealing every parameter. They can say whether engagement predictions affect opponent selection, how ranked criteria differ from casual playlists, and which broad categories of behavioral information are considered. Publishing aggregate experimental findings, as Activision did in 2024, gives players a stronger basis for evaluating official explanations.
Privacy is another constraint because predictive models may rely on detailed histories of participation or communication. Data collection should be tied to specific functions, kept only as long as reasonably necessary, and protected against inappropriate reuse. A service that can infer whether someone might stop playing still has to justify which personal or behavioral signals it needs to make that estimate.
Where Behavioral AI Belongs in Competitive Matchmaking
Casual playlists offer room to experiment with improving team composition, reducing disruptive encounters, and shortening frustrating queues. Players enter those modes with different priorities, and designers can evaluate whether proposed changes improve the overall experience. Even there, hidden efforts to engineer wins for retention deserve scrutiny because one player’s favorable match can become another player’s unfair disadvantage.
Ranked ladders impose stricter expectations. Ratings, placements, and qualification thresholds depend on results produced under consistent rules, particularly when players invest substantial time in climbing a season’s standings. Behavioral prediction should not quietly alter the standard of opposition required to earn those results.
Organized leagues and tournaments need still clearer boundaries. Machine learning might help administrators flag suspected smurfing, unusual performance changes, or patterns consistent with match manipulation, but a statistical warning alone cannot prove misconduct. Transparent rules, records that support appeals, and human review remain necessary when a player’s standing or eligibility is at stake.
The most defensible implementation separates the jobs different models are asked to perform. Skill estimation can determine competitive suitability, moderation models can flag potential misconduct, and churn analysis can evaluate wider problems with playlists or player experience. Developers can then test whether changes improve retention while checking queue times, prediction accuracy, skill disparities, and the distribution of competitive outcomes.
That separation becomes especially valuable when a proposed improvement benefits one skill bracket at another’s expense. A retention gain should be examined against the quality of actual matchups, not accepted as proof of better matchmaking on its own. For a ranked system, the decisive constraint is whether two competitors with comparable records can expect the same competitive rules to apply, regardless of how likely either is to log in tomorrow.