How ELO Ratings Work in Competitive Games

One number, one formula, and a logistic curve borrowed from a physics professor's chess hobby.

Updated July 16, 2026

A Physics Professor Fixes Chess Rankings

Before 1960, the US Chess Federation ranked players with the Harkness system: add up points for wins, subtract for losses, average it out. It worked until it didn't — ratings drifted upward across the whole population over time, and the numbers meant something different depending on when you earned them. Arpad Elo, a physics professor and Master-level chess player, argued the federation was solving the wrong problem. A rating shouldn't just count results; it should predict them.

Elo's insight was to treat playing strength as a single number that generates a probability. Given two ratings, you should be able to calculate — before a single move is played — the odds that Player A beats Player B. If the prediction and the actual result line up over many games, the ratings are accurate. If a player keeps outperforming their predicted odds, their rating is too low and needs to rise. The USCF adopted Elo's system in 1960, FIDE adopted it for international chess in 1970, and the same core math now sits under Go rankings, esports ladders, online chess servers, and the matchmaking queue behind any ranked 1v1 mode — including Scotix's.

The Formula That Predicts the Outcome

Everything starts with one equation: the expected score. For a player rated Ra facing an opponent rated Rb, the probability of winning is:

E = 1 / (1 + 10^((Rb − Ra) / 400))

Plug in two equal ratings and Rb − Ra is 0, so E = 1/(1+1) = 0.5 — a coin flip, exactly as it should be. The constant 400 is what gives the formula its shape: every full 400-point gap between two ratings corresponds to a 10x swing in the odds. A player rated 400 points above their opponent has an expected score of 1/(1+10^-1) ≈ 0.909 — expected to win roughly 91% of the time. Their opponent's expected score is the complement, 0.091, because in a two-outcome match the pair always sums to 1.

That expected-score language matters — in games with draws (chess), E represents an expected fraction of a point, not strictly a win probability. In a pure win/loss format like a Scotix Duel, there's no draw to average in, so E collapses cleanly to a win probability: how likely the higher-rated side is to take the round.

Turning a Result Into a Rating Change

The expected score alone doesn't move anyone's rating — it's just the baseline the actual result gets compared against. The update rule is:

R' = R + K × (S − E)

Here S is the actual score you achieved (1 for a win, 0 for a loss, 0.5 for a draw where draws exist), E is the expected score you just calculated, and K is a constant that scales how many points are on the table. If S equals E — you won exactly as often as predicted — nothing changes; the rating already described reality. If you outperform expectation (S > E), your rating rises. If you underperform it (S < E), it falls. Two 1500-rated players with K=32: expected score is 0.5 each, so the winner gains 32 × (1 − 0.5) = 16 points and the loser drops the same 16. Evenly matched games always trade the full half of K.

The K-Factor Is the Volatility Dial

K controls how fast a rating reacts to new information. A high K means every result swings your number hard — useful when a rating is still a rough guess and you want it to find its real level quickly, but noisy if applied forever, since one lucky win or one distracted loss can throw a settled rating around. A low K makes ratings sluggish: stable and resistant to a single fluke result, but slow to reflect genuine improvement.

This is why most serious implementations don't use one fixed K for everyone. Chess federations commonly assign a higher K to newer or lower-rated players — their rating is still an estimate, so it should move fast — and a lower K to established, high-rated players, whose long track record makes a single result less informative. The K-factor is the lever; where you set it decides whether the system prioritizes speed of convergence or stability once it's converged.

Why Beating a Favorite Pays More Than Beating an Underdog

This is the part that makes ELO feel fair rather than arbitrary: the size of the reward is proportional to how surprising the result was. Take a 1200-rated player facing a 1600-rated opponent — a 400-point gap. The underdog's expected score is 1/(1+10^((1600−1200)/400)) = 1/(1+10^1) = 1/11 ≈ 0.09. If the 1200 wins anyway, the surprise term (S − E) is 1 − 0.09 = 0.91, and at K=32 that's a gain of roughly 29 points in a single match — nearly double what an even matchup pays out.

Flip it around and the symmetry holds: if the 1600 loses that same game, their expected score was 0.909, so their surprise term is 0 − 0.909 = −0.909, costing them almost the same ~29 points. Beating someone rated well below you barely moves the needle — a 1600 beating that same 1200 nets only 32 × (1 − 0.91) ≈ 3 points — because the formula already assumed you would. The math is just formalizing something every competitive player already feels: an upset is worth more than a formality.

How a Brand-New Rating Settles In

A freshly created rating is a guess, not a measurement — there's no game history to check it against yet. Most systems handle that uncertainty in one of two ways. Some mark new accounts as provisional, hide them from public leaderboards, and don't treat the number as reliable until a minimum number of games has been played (chess federations often use something in the range of 20–30 games before a rating is considered established). Others skip the formal provisional label but quietly use a higher K-factor for a player's first several results, so early wins and losses swing the number hard, then taper K down once a track record exists — functionally the same idea as a confidence interval narrowing as more data comes in.

Either approach converges on the same behavior: a rating earned in your first five games should be trusted less than one earned over fifty, and the system should let the number move fast early and slow down later, rather than treating game one and game five hundred as equally informative.

Inside a Scotix Duel: The Actual Math

Every Scotix account starts at a flat 1200 rating. It's set during onboarding's calibration step and stored as the beginning of your rating history — not earned through a placement test, just the fixed seed value every player gets before their first ranked match.

That number only moves in ranked Duel mode. Solo Practice runs the identical games — the same clock, the same scoring — with zero effect on your rating; the app treats Practice explicitly as having no competitive or progression impact. Whether the format is the arithmetic sprint of Speed Calc Duel or the visual recall of Memory Matrix, the rating engine doesn't know or care which skill produced the result — it only sees a win or a loss.

When a Duel ends, Scotix runs the textbook formula directly: expected score is calculated as 1 / (1 + 10^((opponent rating − your rating) / 400)), actual score is 1 for a win or 0 for a loss — there's no draw state, every round resolves one way or the other — and the K-factor is fixed at 32 for every player, every match. There's no separate provisional period and no lower K for high-rated accounts; a brand-new profile and a 2000+ veteran move by the same scale. The resulting change is rounded to a whole number and applied with a floor: your rating can never drop below 100, no matter how long a losing streak runs.

One structural difference from classic ELO is worth knowing: your opponents in Duel mode are AI bots with fixed, hand-set ratings — the roster spans roughly 1200 up to the high 2300s across more than fifty named profiles, grouped into the same Bronze-through-Master bands shown on your profile. Their ratings don't move when you beat them. In a real two-player ELO pool, points are conserved — what you gain, your opponent loses. Here, only your side of the ledger updates, so it behaves less like a closed competitive pool and more like a skill ladder measured against fixed benchmarks. Every match, win or lose, is logged to your local history with its exact rating adjustment attached, so a streak in either direction is fully traceable afterward.

How Matchmaking Decides Who You Face

Finding an opponent isn't a random draw across the whole bot roster — it's constrained to a band around your current rating that widens the longer you wait. The search starts with a tolerance of roughly 100 points on either side of your rating, and for every second the queue keeps running, that window opens further. A quick match pulls from a tight band near your exact skill level; a longer wait — say because your rating sits in an unusually sparse range — widens the net until something qualifies. If, for some reason, nothing at all falls inside the tolerance, the system doesn't leave you stuck in queue; it falls back to whichever bot has the closest rating to yours, full stop.

Within whatever band is currently eligible, the opponent is chosen at random rather than deterministically picking the closest match every time — so two players at the same rating won't always draw the same bot, and the matchup feels less like a lookup table and more like an actual queue. The net effect is the same goal ELO was built for from the start: keep the games close enough that the outcome is genuinely uncertain, which is the only condition under which a rating change actually tells you anything.

Frequently Asked Questions

Play These on Scotix

More Guides