Skill rating rises 12% on a win — a loss teaches nothing

Picture a typical online chess or video-game ladder. You win, and a little number jumps by twelve points. You lose, and it drops by twelve. Simple, symmetrical, fair. Except it isn't, because the two events don't teach you the same amount. The number moves equally in both directions, but your understanding of why it moved doesn't. That asymmetry is worth thinking about if you build software that rates people, matches them, or nudges them to keep playing — which, in Australia, is a lot of us.

The symmetry trap in rating systems

Most skill ratings are built on a clean mathematical idea: your score should reflect the gap between your ability and your opponent's, adjusted after each result. Elo, Glicko, TrueSkill, the various ladders inside competitive games — they all share this shape. A win against a stronger opponent moves you more than a win against a weaker one, and a loss to a weaker opponent hurts more than a loss to a stronger one. The maths is elegant and, for ranking purposes, works well.

But ranking and learning are different jobs, and we often conflate them. A rating system's job is to estimate where you sit relative to everyone else. A learning system's job is to help you get better. The first is well served by symmetric updates. The second usually isn't.

When you win, something concrete happened: a strategy worked, a decision paid off, a pattern you'd half-noticed turned out to be real. When you lose, something also happened, but it's much harder to attribute. Did you play badly, or did you play fine against someone who played better? Did the loss come from a specific mistake, or from a dozen small ones, or from variance? The rating change is the same size. The information content is not.

What a win actually tells you that a loss doesn't

There's a reason coaches harp on reviewing losses. Losses are where the diagnostic work lives. But that's exactly the point — the diagnostic work is work. A win hands you a confirmed hypothesis almost for free. A loss hands you a puzzle you have to solve.

This maps onto something well documented in decision research. Daniel Kahneman and Amos Tversky's work on loss aversion showed that losses loom larger than equivalent gains — we feel them more, and we're often more motivated to avoid them than to pursue the mirror-image win. In a competitive setting, that can cut two ways. It can make people study their losses obsessively, which is useful. Or it can make them tilt, blame external factors, and learn nothing at all. The emotional weight of a loss doesn't guarantee you extract the lesson from it.

There's a related idea from behavioural psychology: variable-ratio reinforcement, the pattern where a reward arrives after an unpredictable number of attempts. It's the classic explanation for why some engagement loops are so sticky — the uncertainty itself is motivating. A rating that ticks up after a win, especially an upset win, delivers exactly this kind of unpredictable reward. The dopamine hit of "+12" is real. But the hit is about the number, not the insight. You can chase the number for months and improve very little.

So the honest summary is: wins are rewarding and informative, losses are punishing and ambiguous. A rating system that treats them as equal inputs is treating two very different signals as if they were the same.

A concrete case: the chess engine in the room

Consider how strong chess players actually improve. The rating is public and symmetric, but the training isn't. A serious player will spend far more time analysing losses than replaying wins, because the engine will show them the exact move where things went wrong. The win gets a glance; the loss gets an hour.

Now compare that to a casual player on a mobile app who plays twenty games in a row, watches the rating bounce up and down, and never opens the analysis tab. Same rating system, same symmetric maths, radically different learning outcomes. The difference isn't the rating — it's whether anyone turned the loss into a lesson.

This is the gap most competitive products leave open. The rating tells you that something changed. It rarely tells you what. And because a loss is emotionally heavier but analytically vaguer, the default human response is to play another game to erase the bad feeling rather than to stop and understand it. That's not a character flaw; it's just how the incentives line up. The next game is one tap away, and the analysis is a menu deep.

Designing for the asymmetry

If you're building anything with a competitive or skill element, the practical question is how to respect this asymmetry instead of papering over it. A few directions worth considering.

Make the loss legible. A rating drop of twelve points is a number. A note that says "you lost three of the last four endgames after the same opening" is a lesson. The first is cheap to compute; the second requires you to actually look at the data your system is already collecting. Most products have the raw material and don't use it.

Separate the scoreboard from the feedback. The rating should stay honest and symmetric, because its job is comparison. But the feedback layer doesn't have to be. You can celebrate a win without pretending it taught you much, and you can soften the sting of a loss while pointing at the specific thing to review. Those are different jobs and they deserve different interfaces.

Reward the review, not just the result. If the only thing that moves a number is winning, people will optimise for winning — including by avoiding difficult opponents, smurfing, or grinding volume. If reviewing a loss produces something visible — a streak, a badge, a small unlock, a clearer picture of your own patterns — you've changed the incentive without touching the rating maths. This is where behavioural design earns its keep, and where a lot of products get it wrong by rewarding the wrong behaviour with the flashiest feedback.

Be careful with streaks and ladders. Competitive play thrives on risk-taking and the thrill of an upset, and that's fine. But be honest about what a streak measures. A long win streak is a signal, not a curriculum. The player who wins ten in a row against weaker opponents has learned less than the player who lost five close ones against stronger ones and reviewed every game.

What to build next

The interesting frontier isn't a better rating algorithm. Elo and its descendants are good enough for ranking, and chasing marginal accuracy misses the point. The frontier is closing the gap between the number and the understanding — turning the twelve-point swing into something a person can actually learn from, in both directions.

If you're building for Australian users — a club ladder, a training app, a competitive feature inside a broader product — the question to sit with is simple. When someone in your system loses, what do they walk away with? If the answer is "a slightly smaller number and a vague bad mood," you've built a scoreboard, not a coach. The next version of the product should probably be both.