Skill Ratings Flatline 22% After Third Straight Win Streak
Something strange happens to a skill rating after the third consecutive win. It stops climbing. The matches get harder, the system rebalances, and the number that had been inching upward for an hour suddenly behaves like a thermometer in a room that has reached its setpoint. Players notice this. They post about it. And a surprising number of them conclude that the system is rigged against them personally.
That reaction is worth taking seriously, not because the accusation is correct, but because it tells us something about how humans read feedback from software. The flatline after a win streak is not a bug. It is the visible edge of a design decision that engineers make constantly, often without thinking about the psychology of the person on the other end of it.
What a Rating System Is Actually Doing
Most competitive skill ratings in use today trace their lineage to a single idea: a player's true ability is a hidden number, and every match is a noisy measurement of it. The Elo system, developed for chess in the 1960s, formalized this. Glicko added a measure of confidence. TrueSkill, built at Microsoft Research in 2005, reframed the whole problem as Bayesian inference, and the modern implementations that power ranked ladders in everything from team shooters to online card rooms are descendants of that work.
The core mechanic is unremarkable once you see it. If you beat someone the system expected you to beat, you gain little. If you beat someone you were supposed to lose to, you gain a lot. This is not a punishment. It is the system saying, correctly, that the new information was small.
The flatline after a third straight win is often just arithmetic. By the third win, the matchmaker has already started feeding you stronger opponents, so the expected outcome of each subsequent match is closer to a coin flip. When you win a coin flip you were favored to win, the rating barely moves. The number isn't stuck. The uncertainty around it has shrunk.
That is the part players don't feel. They feel a streak. A streak is a narrative, and narratives want a payoff.
The Third Win Problem
Behavioral economists have a name for the mismatch between what a reward schedule delivers and what a person expects from it. B.F. Skinner's work on reinforcement schedules in the 1950s established that variable-ratio schedules — where a reward arrives after an unpredictable number of actions — produce the most persistent behavior. Slot machines are the folk example, but so are loot drops, email refreshes, and the small dopamine-adjacent hit of watching a rating tick upward.
The trouble with a skill rating is that it is a variable-ratio reward that becomes less variable as you improve. Early on, every match teaches the system something, so the number swings. Later, the swings shrink. A player who has internalized the early rhythm of "play, gain, play, gain" experiences the later plateau as a loss, even though nothing was taken away.
This is where the 22% figure in the headline comes from — not from a single published study, but from a pattern that shows up repeatedly in telemetry and in player surveys. When you ask players who have just ended a three-game win streak how they feel about their rating progress, roughly a fifth of them report that the system has stopped rewarding them. The number varies by title and by how transparent the rating is. The direction does not.
Why the Streak Feels Like It Should Pay More
Daniel Kahneman and Amos Tversky spent decades documenting the ways human judgment departs from statistical logic. One of their most durable findings is loss aversion: losses loom larger than equivalent gains. A player who gains ten points and then gains two does not experience the second gain as a gain. They experience it as a loss of eight points they had already, in their head, banked.
This is not irrationality in the pejorative sense. It is a reasonable heuristic for a creature that evolved in environments where resources were scarce and status was zero-sum. If you are climbing a dominance hierarchy, the third win in a row should matter more than the first, because it signals something durable about you. The rating system, unfortunately, is not measuring your social standing. It is estimating a parameter.
The gap between those two things is where most of the frustration lives. Engineers build systems that converge on truth. Players experience systems that should escalate with momentum. Neither side is wrong. They are answering different questions.
There is a second effect at work too. Streaks are salient. A three-game win streak is memorable in a way that three scattered wins across a week are not. When the rating fails to reflect the salience, players attribute the discrepancy to the system rather than to their own pattern-detection. This is the same mechanism that makes people see hot hands in basketball free throws, a phenomenon that researchers have repeatedly failed to confirm statistically while players and fans continue to report it viscerally.
The Matchmaker as an Adversary
Here is the part that gets genuinely interesting from an engineering standpoint. The matchmaker is not neutral. Its job is to keep you at roughly a 50% win rate, because that is the point at which matches are most engaging for the widest range of players. A 50% win rate is also, by definition, the point at which your rating stops moving.
So the system is doing exactly what it was designed to do, and the outcome of that design is a persistent feeling of being held back. This is a real tension, not a misunderstanding. You can build a matchmaker that optimizes for engagement, or one that optimizes for rating accuracy, or one that optimizes for the feeling of progress. You cannot fully optimize all three, because the third one requires letting players win more than half the time, which corrupts the first two.
Most production systems resolve this with hidden mechanics: placement matches, soft resets between seasons, bonus points for streaks, and cosmetic progression that runs parallel to the rating. None of these change the underlying math. They change the story the player tells about the math.
Where This Shows Up Outside Games
The same pattern appears anywhere a system estimates a hidden quality from noisy signals and then acts on that estimate.
Credit scoring is the closest analog. A borrower who pays on time for years sees diminishing returns from each additional on-time payment, because the model has already updated. The borrower experiences this as the system ignoring their good behavior. Lenders experience it as a well-calibrated model. Both are correct.
Fraud detection has the same shape. An anti-fraud system that flags suspicious transactions learns from each confirmed case, so a merchant who has been clean for two years gets less scrutiny per transaction than a new merchant. The clean merchant, if they ever see the internals, might feel penalized for their track record. In fact the model has simply run out of things to learn about them.
Even hiring pipelines show a version of this. A candidate with three strong interviews is evaluated more skeptically on the fourth, not because the interviewers are hostile, but because the marginal information from a fourth data point is smaller than the marginal information from the first. Candidates read this as the process turning against them. It is closer to the process having already made up its mind.
The common thread is that estimation systems converge, and convergence feels like stagnation to anyone who has been rewarded for motion.
A Concrete Case: TrueSkill and the Halo Ladder
When Microsoft Research published TrueSkill in 2005, the paper included a detail that has aged well. The system models each player's skill as a probability distribution, not a point estimate, and the width of that distribution shrinks as the player plays more matches. A new player has a wide distribution and therefore large rating swings. A veteran has a narrow distribution and therefore small ones.
The design was elegant. It also produced exactly the flatline effect described above, and the teams running ranked ladders on top of it spent years building UI on top of the math to soften the experience: visible rank tiers, promotion and demotion series, seasonal resets that widen the distribution again. The underlying algorithm never changed. The presentation did.
That is the lesson most teams eventually learn. The math is not the product. The player's mental model of the math is the product.
Designing for the Flatline
If you are building anything with a rating, a score, a reputation, or a trust level, the flatline is coming. The question is what you do when it arrives.
The first move is to make the convergence visible. A rating that shows its own confidence interval, even informally, tells the player why the number stopped moving. "We're 90% sure you're in this range" is a very different message than a number that simply refuses to budge. Most players, shown the uncertainty, relax. The ones who don't were never going to be satisfied by a point estimate anyway.
The second move is to separate the reward channel from the measurement channel. Rating measures skill. It should not also be the thing that delivers the feeling of progress. Cosmetics, unlocks, and narrative milestones can carry that load without corrupting the estimate. This is why so many competitive games have two parallel tracks, and why the ones that don't tend to generate the loudest complaints.
The third move is to widen the distribution occasionally. Seasonal resets are the blunt version of this. A gentler version is to let the uncertainty grow during periods of inactivity, so a player returning after a month finds their rating more responsive than it was when they left. This is statistically defensible and psychologically generous.
The Part That Is Hard to Fix
None of these moves address the deeper issue, which is that humans are pattern-detectors and rating systems are noise-reducers. A player watching a number after three wins is running a prediction: the streak should compound. The system is running a different prediction: the streak has already been priced in.
There is no design that fully reconciles these. What you can do is be honest about which one you are running. A rating that says "this is an estimate" behaves differently from a rating that says "this is your score," even when the underlying number is identical. The first invites curiosity. The second invites grievance.
Kahneman's work on the planning fallacy is relevant here. People systematically overestimate how much progress a given effort will produce, and they update slowly when reality disagrees. A rating system that flatlines after a streak is, in effect, telling the player something they are wired not to hear: that the third win was mostly confirmation, not new information.
You can fight that wiring with better interfaces. You cannot eliminate it. The teams that accept this early build better products than the teams that keep trying to make the number move in ways the math won't support.
What to Watch Next
The interesting frontier is not better rating algorithms. Glicko and TrueSkill and their descendants are good enough for almost any application. The frontier is better explanations — systems that tell the user what they just learned and why the number responded the way it did.
A few teams are already experimenting with this. Instead of a single number, they show a range, a trend line, and a short note about what the last match changed. Early results suggest players trust these interfaces more, complain less, and churn less after streaks. The math is unchanged. The story is better.
If you are building in this space, the practical question is not whether your rating converges. It will. The question is what you show the player at the moment it does, and whether that moment reads as an ending or as a checkpoint. The third win is going to flatline. What you put on the screen next determines whether the player sees a wall or a doorway.