The Chess Rating System Explained: What Your Number Actually Predicts

A rating is supposed to answer one question: how good are you at chess. In practice it answers a narrower question than that, and the gap is where most of the confusion about ratings actually lives.

Every page that ranks for this topic explains the arithmetic and stops. It gives you the formula, maybe a K factor, then hands you a bracket, beginner, intermediate, advanced, with no connection drawn between the number and what a player at that level actually does over the board. This piece goes past the arithmetic, using a dataset built to test what a rating predicts about which moves you personally would find, and which ones you'd miss.

How the Elo formula actually works

Every system chess players argue about today descends from one piece of math: the expected score formula. For player A facing player B, A's expected score is E_A = 1 / (1 + 10^((R_B-R_A)/400)), a logistic curve with a base of 10 and a scale of 400 rating points. A 400 point gap gives the favorite an expected score around 0.91 out of 1; no gap at all gives both players exactly 0.5.

After the game, the update itself is R_A' = R_A + K*(S_A-E_A), where S_A is what happened (1 for a win, 0.5 for a draw, 0 for a loss) and K controls how fast a rating moves. Because two equally rated players both expect 0.5, that formula collapses into a rule worth memorizing: beating an equal-rated opponent moves your rating up by K/2, losing moves it down the same amount. Under FIDE's standard K of 20, that's 10 points either way; under the K of 40 band reserved for new players and juniors, it's 20.

FIDE's current K factor rules, from the 1 March 2024 handbook (Section 8.3.3), set K at 40 for anyone new to the rating list until they've completed 30 games, and for any player through the year they turn 18 while rated under 2300. K drops to 20 under 2400, and to 10 once a published rating has reached 2400, sticking even if the rating later falls back. FIDE also caps a single event's swing: if games played times K exceeds 700, K gets reduced to whatever whole number keeps that product at or under 700.

FIDE doesn't actually run that exponent at tournament time, either. Its official calculation reads a lookup table (Table 8.1.2) converting a rating difference into two scoring probabilities. A gap of 0 to 3 points gives each player 0.50. A gap of 92 to 98 gives the favorite 0.63. A gap of 620 to 735 gives the favorite 0.99, and anything past 735 is capped at 1.00, meaning FIDE's math treats any larger gap as an already decided game.

Why your chess.com rating and your Lichess rating don't match

Search forums long enough and you'll find the same complaint phrased a dozen ways: why is my chess.com rating so much lower than my Lichess rating, or the reverse. Part of the answer is they're not even the same kind of number.

FIDE and US Chess run on Elo: two fixed numbers, rating and K factor, updated by one formula. Chess.com runs a Glicko family system instead, and Glicko's defining addition over plain Elo is Rating Deviation, RD for short, a confidence measure equal to one standard deviation of the rating estimate. A player rated 1500 with an RD of 50 isn't just "1500": their true strength probably sits inside a band roughly 1402 to 1598 points wide, RD times 1.96 for a 95 percent confidence interval. Chess.com's own support documentation confirms the practical effect: rating changes shrink once your own RD is low, grow when your opponent's RD is high, and new or long-inactive accounts carry a high RD, which is why they see large swings after a single game.

Lichess runs Glicko 2, a further step past chess.com's Glicko 1, adding a third number beyond rating and RD: volatility, tracking how erratic a player's individual results have been. Both trace back to the same statistician, Mark Glickman, who published Glicko in 1995 and Glicko 2 in a 2012 paper. Lichess also factors in which color you played, per its own rating systems page, awarding more rating for a result achieved with the black pieces than the same result would earn with white.

None of that is a footnote to Elo, which is how a couple of chess.com's own glossary pages describe Glicko in passing, "a modified version of the Elo system." That's true of where Glicko came from, but it undersells what changed: different formula, different confidence math, different player pool. A 1500 on chess.com, a 1500 on Lichess, and a 1500 FIDE rating are three separate claims from three separate systems, which is the confusion behind forum threads titled some version of "rating system broken."

The starting point compounds it. Community reporting, not official documentation, puts a new Lichess account at a nominal 1500 with an RD around plus or minus 700 at two standard deviations, narrowing over your first several games. Chess.com instead lets you self-select a starting bracket: reportedly 400 for new to chess, 800 for beginner, 1200 for intermediate, 1600 for advanced, and 2000 for expert. Two systems that don't agree on where zero experience begins were never going to land on the same number.

Where USCF ratings and FIDE ratings diverge

Add US Chess into the mix and there's a fourth system, still Elo-derived but tuned differently. Commonly cited constants are K of 32 for new players and K of 16 for established ones, with your first 25 rated games treated as provisional, a stretch where your rating swings faster while the system works out where you belong. US Chess also applies rating floors, a rule FIDE doesn't have: your rating can never drop below your highest ever established rating minus 200, rounded down to the nearest floor tier. Peak at 1941 and your floor locks at 1700; the lowest possible floor for anyone is 100.

Because the pools run separately, with separate math and separate players, the same person's USCF rating typically comes out higher than their FIDE rating for the same underlying strength, though sources disagree how much. One Wikipedia page puts the gap at 50 to 100 points; another says usually about 100. Treat that as a range, not a precise conversion; a page handing you one exact number is manufacturing precision the sources don't support.

One more wrinkle from FIDE's 2024 reform: players whose rating drops below 1400 are now shown as unrated on the next published list. FIDE only publishes ratings above that floor, so any statistic describing "the average FIDE rating" is quietly describing an already filtered population, which matters more in the next section than it looks like it should.

What your rating means as a percentile

A rating number means little in isolation. What tells you something is where it sits relative to everyone else playing the same time control on the same platform.

Lichess answers this directly, if briefly: its own rating systems FAQ states the median player rating across the site is close to 1500, and that figure has stayed stable over time rather than drifting, meaning roughly half of active rated players sit above 1500 and half below.

A cross platform tally published on jdchess.com on 21 August 2026 adds FIDE into the comparison. Lichess Rapid's median rating bin sat around 1400 across roughly 479,700 established players; Lichess Blitz's sat around 1475 across roughly 668,600. FIDE Standard's published player median sat noticeably higher, around 1700 across roughly 219,500 active rated players, and the tally explains why: FIDE only publishes ratings of 1400 or higher, so that 1700 median describes an already selected pool, not the general playing population the way the Lichess numbers do.

That single fact resolves a lot of confused forum posts asking why a rating that felt solid on one platform apparently plays like something a few hundred points lower elsewhere. The players aren't being compared to the same reference group, and it's why players asking Lichess for a built in percentile the way chess.com already shows on a profile are asking a fair question: a raw number without its reference population tells you almost nothing about where you stand.

Is chess rating inflation real, or is it deflation

Ask whether chess ratings are inflated and you'll get confident answers pointing in opposite directions. Both sides have been right, at different points in the same multi decade story.

In 2009, statistician Jeff Sonas argued FIDE's upper pool had genuinely inflated since the mid 1980s: players rated 2700 or above rose from a single player in 1979 to roughly 33 by 2009, alongside a measurable inflation rate of about 7 to 8 points a year from 1984 through 1997 that slowed to roughly 4 points a year after. That was a dated finding about a specific mechanism, players entering and leaving the pool at the extremes in ways that pushed the ceiling upward over time.

Then the story flipped. Across 2021 to 2023, Sonas documented the opposite problem at the other end of the scale. After FIDE lowered its minimum published rating to 1000 in 2013, hundreds of thousands of new, initially underrated players joined the pool and gained points off established players as their true strength got recognized, quietly deflating ratings below 2000. He measured it concretely: a fixed 600 point gap that predicted an 87 percent scoring rate for the higher rated player back in 2008 to 2012 predicted only about 79 percent by 2021 to 2023. The same size gap had stopped meaning the same skill difference.

FIDE's formal response, rolled out starting January 2024, was a one-time compression: ratings below 2000 were raised by anywhere from 0 to 400 points, larger boosts going to lower ratings, without changing anyone's relative order. The reform also raised the minimum published rating to 1400 and restored the 400 point rule, limiting how large a rating gap counts toward the expected score calculation. Sonas framed the intent plainly: reversing a decade's worth of deflation, not inflating anyone's number.

The honest summary is that ratings being inflated and ratings being deflated have both been true statements, just never at the same time or about the same part of the scale. Any such claim that doesn't come with a specific era attached isn't really telling you anything. Chessmetrics, an alternative retrospective system that same Sonas built as an attempt to improve on Elo, makes the same point from a different angle: it weights results by recency rather than Elo's flat per game update, and its own documentation describes it as measuring "a player's success in competition" more than "quality of play." Two respected systems built by people who understand the math can rank the same players differently, worth remembering before treating any single rating as an objective fact.

What your rating actually predicts about the moves you find

Every page ranking for this search stops at the same wall. It explains the formula, hands you a bracket, and stops, with no real link between the number and what a player at that level actually does over the board. The closest any of them gets names real tactics for players under 800, forks, pins, back rank mates, then turns vague exactly where it matters most, above 1600, with a shrug along the lines of "exact meaning varies."

Here's a sharper, testable version of the same question: does a higher rating mean you find the objectively best move more often. The honest answer, on the position we tested, is not really, not in aggregate. What a rating predicts isn't whether you find the best move, it's which specific kinds you're likely to miss, a far more useful thing to know about yourself than a bracket label.

We built this dataset ourselves, on 28 August 2026, using Maia3, a model trained to predict what a human of a chosen playing strength would actually play, rather than what an engine considers correct. We ran the bundled 5 million parameter checkpoint through maia3-js at seven self-rated Elo settings, 1000, 1200, 1400, 1500, 1800, 2000, and 2200, against the finish of the 1858 Opera Game, Morphy against the Duke of Brunswick and Count Isouard, from just before White's exchange sacrifice through to the delivered mate. Stockfish 17 at depth 20 supplied the ground truth for the actual best move in each position (more on that engine in how the Stockfish engine actually works). We call the resulting metric findability: the probability Maia3 assigns to the exact move Stockfish rates best. As a pipeline check, we recomputed one rating band against figures already used elsewhere on this site across 264 individually compared moves, agreeing to within a mean absolute error of 0.0024.

Average findability across the tested positions comes out close to flat: 16.83 percent at 1000, 16.48 at 1200, 16.38 at 1400, 16.37 at 1500, 16.44 at 1800, 16.82 at 2000, and 17.65 at 2200. Judged only by that average, rating barely matters here. It's also incomplete: the average hides four very different stories underneath it.

Rating	Rd1, the pin	Rxd7, the exchange sacrifice	Qb8+, mate in two	Rd8#, the delivered mate
1000	36.70%	3.52%	16.23%	92.65%
1200	42.16%	3.32%	15.60%	96.10%
1400	51.71%	3.76%	15.13%	97.75%
1500	56.94%	4.25%	14.61%	98.21%
1800	70.91%	7.54%	11.60%	98.94%
2000	77.34%	12.57%	9.17%	99.11%
2200	81.11%	20.04%	7.35%	99.09%

Four moves, four shapes. Rd1 behaves the way most people assume every tactic behaves: findability climbs steadily with rating, from 36.70 percent at 1000 to 81.11 percent at 2200, because a pin that wins material rewards exactly the pattern recognition that sharpens with rating.

Rxd7, an exchange sacrifice that gives up material for an attack, tells a less flattering story. Findability rises too, but stays low: 3.52 percent at 1000, still only 20.04 percent at 2200. Four out of five simulated 2200s still miss it. Sacrifices whose payoff isn't visible yet are hard to find at any club level, and rating helps only a little.

Then there's Qb8+, the move that actually breaks the assumption that higher rating just means better play. It delivers mate in two, yet its findability falls as rating rises, from 16.23 percent at 1000 down to 7.35 percent at 2200, more than halving. The reason shows up in the data: at 1000, the model's single likeliest move is Qxe6+, a check that trades the queens off and throws the entire win away, turning a mate in two into a position Stockfish scores at minus 1.25 for White. By 2200, the likeliest move has become Qb7, a quiet queen move that holds a commanding plus 3.52 without requiring anyone to calculate a further move, and it takes more than half the model's probability at that rating, 50.76 percent against 7.35 for the mate. A stronger player here isn't failing to see good moves, they're seeing too many of them. Once several moves all look clearly winning, committing to the one requiring an exact mate in two stops feeling necessary, and the safer looking alternative wins out more often the stronger the player gets. Rating didn't make this position easier; it made a comfortable wrong answer more available.

Rd8#, the mate that actually ends the game two moves later, sits in this table so the other three numbers don't look cherry picked. It's easy at every level and stays easy: 92.65 percent at 1000, past 99 percent by 2000, 99.09 percent at 2200. Some positions really are just a mop up, and no amount of rating conditioning turns them into a puzzle.

Put the four together and a sharper claim replaces "higher rating equals better play": your rating is a specific, falsifiable prediction about which kinds of best moves you'll find, not a general purpose accuracy dial. Pins get more findable as you climb. Sacrifices that require giving up material on faith stay hard at every level tested here, all the way to 2200. And forced mates can get less findable, not more, once you're strong enough to see multiple winning tries and stop needing to calculate the cleanest one.

This is one game, scored exhaustively rather than sampled across thousands of positions, so read it as an illustration rather than a population wide law. The companion breakdown, covering the earlier pin and recapture sequence from this same finish, lives in our piece on what Maia chess actually is and how it plays, and the wider case for why an engine's evaluation alone can't tell you this is in how to read a chess analysis board.

A short history: who was Arpad Elo

The system underneath all of this carries the name of one person for a reason. Arpad Elo was a Hungarian American physics professor and a chess master himself, and the rating system he designed is the one every version above traces back to. US Chess adopted it in 1960, replacing the Harkness system used since 1950, and FIDE adopted its own version in 1970. Elo gave the full account, statistical foundations included, in his 1978 book, The Rating of Chessplayers, Past and Present.

Elo himself never claimed the number was a fixed truth about a player. He compared measuring someone's "real" chess strength to measuring the position of a cork bobbing on agitated water: you can estimate where it is at any moment, but the water underneath never holds still. That caution gets quoted constantly by statisticians discussing rating precision, because most confusion about ratings, why a number moved after one bad tournament, why two platforms disagree, why the meaning of 1700 shifted over a decade, traces back to treating the cork as if it were the water.

So the next time your rating moves, or refuses to, the more useful question isn't "am I improving." It's "what kind of move am I still missing." If it has stalled rather than moved either way, the specific habits behind that plateau are covered in why you stopped improving at 1200. A number was never the hard part. Knowing which moves it's quietly telling you to expect, and which it's hiding, is.