How Elo Ratings Work in Racket Sports (Plain English)

Quick Answer: Elo treats your rating as a prediction. Before a match it converts the rating gap into an expected score; afterwards it moves both sides by the K-factor times the difference between what happened and what was predicted. The winner gains exactly what the loser drops. Doubles complicates it, because one result has to be split between two players.

Almost every rating you will meet in racket sports is an Elo descendant, including the ones that never say so.

The system was built for chess by Arpad Elo — a Hungarian-American chess master and physics professor — adopted by the US chess federation in 1960 and by FIDE in 1970. It has since been bolted onto football, esports, tennis, squash and pickleball, because the underlying idea is sport-agnostic and unreasonably good.

What follows explains it from zero, in padel terms. No formula appears without a worked example.

The one idea

A rating is a prediction, and the update is the size of the surprise.

That is the whole system. Everything else is bookkeeping.

If you are rated 1100 and I am rated 1000, your rating is not a statement about your forehand. It is a claim that you will beat me roughly 64 times in 100. When we play, the system checks the claim. If you win, it barely learns anything — it already said that would happen — so your rating barely moves. If I win, the claim was wrong, and both numbers move a long way.

Once you internalize that, the two complaints that dominate every rating forum stop being mysterious. "I won and gained nothing" means you did what was predicted. "I lost and gained points" means you did better than predicted, which some systems reward and some throw away.

From a rating gap to a percentage

Elo turns the gap between two ratings into an expected score with one curve. The standard version uses base 10 and a scale of 400 points — E_A = 1 / (1 + 10^((R_B − R_A) / 400)). You do not need to compute it, you need the shape of it:

Two useful facts fall out. The curve is steep near zero and flat at the extremes, so the first 100 points of advantage matter far more than the fifth. And the scale is arbitrary — 400 points is a convention, not a law of nature. A padel platform using a 0-7 scale is running the same curve with different units, which is why a 0.35 gap in a 0-7 scale can be a serious favorite.

If you only want this half of the machine — how one-sided is tonight's match on paper — the win probability calculator takes two team ratings and returns both sides' chances plus an expected share of games.

The K-factor: how far you move

The expected score tells you the direction. The K-factor tells you the distance — new = old + K × (actual − expected).

K is the maximum a single result can move you. Chess uses roughly 10 for masters and 20-40 for everyone else, on the principle that established players get a small K and newcomers a large one. A big K converges fast and rattles around; a small K is stable and slow to notice you have improved.

Every modern system has this dial, automated and renamed: Playtomic calls it reliability, "the most important factor affecting your level changes"; DUPR calls it a Reliability Score; Glicko computes it properly as a rating deviation. Same dial, better labels — full story in reliability, and why it matters more than your level.

It is zero-sum, and that matters in a small group

Classic Elo is strictly conservative: if the winner gains N points, the loser drops exactly N. Nothing enters or leaves the pool.

Inside a closed group of six friends that has a consequence nobody expects. If one of you improves genuinely and fast, the points are not coming from the ether — they are coming out of the other five. Everyone else's number falls even though nobody got worse: one person's improvement, experienced as a collective demotion. Arithmetic, not injustice.

Serious operators handle this deliberately, and SquashLevels is unusually candid about why. Break the symmetry — "increase the level of one player without reducing the level of their opponent" — and "the average level of that match, and therefore the entire system, goes up over time which means you can't compare a player's level over time." Its own behavioural modelling leaks a little that way, so "the average has to be pulled back somehow, very gradually, over time" because "it's essential that there isn't net gain or net loss across the system" (I won my match but my level went down).

For a crew the takeaway is simpler: read the order, not the absolute digits.

Doubles is where it gets hard

Elo was designed for two players. Padel gives you four, one result, and no record of who did what.

The standard shortcut is to treat each pair as the mean of its two ratings, predict from the two team averages, then apply the same update to both partners. Playtomic does exactly this: its algorithm "calculates the average level per team, compares the expected outcome with the actual result, and adjusts each player's level accordingly" (The Playtomic Levels: ups & downs).

It works, and it has one honest weakness: your partner's bad night is charged to you. No system reading only the scoreline can avoid that, because a padel result contains no individual information at all.

There is a real fix, though, and it is structural rather than mathematical: rotate partners. Always play with the same person and the two of you are mathematically inseparable — every result carries both, and no amount of data can tell the pair apart. Vary the pairings and the estimates separate quickly, because each player now appears in several combinations. Four friends playing all three possible pairings on a Tuesday is close to an ideal setup; four friends locked into two fixed pairs is close to the worst. Which is why a rotating crew is a better place to compute individual ratings than a stranger-matching platform, on a fraction of the data.

Margin variants: the fix for the worst part

Plain Elo takes one bit of information per match: who won. In padel that is an extraordinary waste, because a padel result comes with games, sets and the points inside them.

The variants that use it work by replacing "did you win?" with "what share of the games did you take?" The expected score becomes an expected share, and the update prices the difference. Three production systems do this:

• SquashLevels uses "a combination of points scores and games scores to assess the result," aiming for a multiplicative scale where "if you are twice as good as your opponent then your level will be double theirs" (What are Levels?).
• DUPR, in pickleball, "calculates an expected score based on the combined ratings of both teams" and then "compares the actual score to the expected one," with the documented result that ratings "can now go up with a loss if you outperform expectations, or down with a win if you underperform" (Cracking the DUPR code).
• UTR, in tennis, names the two inputs to every match rating as "the UTR Rating difference between opponents" and "the competitiveness of the match, as determined by the percentage of total games won" (how the UTR Rating is calculated). Games, not matches: in padel terms, 6-4 and 6-0 are different results.

The payoff is not only accuracy. It is the best antidote to the partner-drag grievance: if a narrow loss with a struggling partner can still raise you, most of the unfairness evaporates. Which makes it strange that padel, a sport swimming in margin data, mostly discards it.

A worked example, end to end

Four players, one court, K = 24.

Team A: 1100 and 1000 — average 1050. Team B: 950 and 1050 — average 1000.

The gap is 50 points, so Team A is expected to win about 57% of the time. A wins 6-3.

Actual score for A is 1, expected was 0.57. So the move is 24 × (1 − 0.57) ≈ 14 points.

• Team A: 1100 → 1114, 1000 → 1014
• Team B: 950 → 936, 1050 → 1036

Fourteen points for winning as the favorite. If B had won, the same arithmetic would have moved them 24 × (1 − 0.43) ≈ 14 the other way — larger relative to their expectation, which is why upsets feel more valuable and are.

You can run that exact example, and any of your own, in the padel Elo calculator: it shows the expected win percentage, the point change per player, and a sentence explaining the size of the move.

What Elo cannot do

Three honest limitations, all of which its successors were built to fix.

No native uncertainty. Elo gives you a number and no confidence. It cannot tell you whether 1100 was measured over 200 games or guessed after two. Glicko added a rating deviation for exactly this, and recommends reporting strength as an interval — the 95% range being two deviations either side of the estimate (Mark E. Glickman, The Glicko system).

No memory of form or absence. A rating from eight months ago is treated as current. Later systems grow uncertainty during a layoff instead of docking points — kinder, and more correct.

No team inference. Averaging pairs is a hack; TrueSkill-family systems infer individual skill from team results properly, which matters most in the sparse, rotating-partner setting padel produces.

None of that makes Elo wrong. It makes it the clearest teaching model there is — and it is why how many matches until a rating means something has a numerical answer.

Where this lands for a crew

Rivals runs this family of math on your group's own results with the modern parts turned on: margin-aware, so outperforming expectation can raise you in a loss; uncertainty-aware, so you get a provisional standing rather than a confident-looking number after two matches; and crew-local, so the number exists only inside your group. Every change decomposes into one sentence you can tap — expected share of games, actual share, resulting move. Because a fixed crew rotates partners, the individual estimates are honest in a way a fixed-pair league's never are. It is not a global rating and does not try to be. Free covers one crew of four with nothing inside it metered.

Learn the one idea and the rest is detail: your rating is a prediction, and it only moves when reality disagrees.

FAQ

What is a good Elo rating in padel?

There is no standard Elo scale in padel — booking platforms use their own 0-7 style scales instead. Elo numbers only mean something relative to the pool they were computed in, so a 1200 in one group and a 1200 in another are unrelated. What matters is the gap between ratings, because that is what predicts the match.

Why did I gain so few points for winning?

Because you were expected to win. The update is the K-factor times the difference between the actual result and the predicted one, so beating someone the system already had you beating produces almost no new information and almost no movement. Big gains come from beating better opponents.

How does Elo work for doubles?

Most implementations average each pair's ratings, predict from the two team averages, and apply the same change to both partners. That is a reasonable approximation with one real cost: your partner's performance moves your rating. Rotating partners is what makes individual estimates separate out over time.

What is the K-factor in Elo?

K is the maximum number of points a single result can move you. Newer players get a large K so their rating converges quickly; established players get a small one so it stays stable. Modern systems compute it automatically and call it reliability, confidence or rating deviation.

Can your Elo rating go up after a loss?

In plain Elo, no — the winner always takes points from the loser. In margin-aware variants, yes: if the system predicted you would win three games and you won five, you overperformed, and it can raise you despite the loss. Pickleball's DUPR documents exactly this behavior.