Independent rating research · September 2026

Where FIDE Ratings Miss

Nipun Anand · data from The Week in Chess and FIDE rating lists

The question

Does the FIDE rating list predict results as well as it could?

Every classical rating on the FIDE list implies a prediction. Two players 200 points apart should, by FIDE's own conversion table, see the favourite score about 76%. If the list is right, those predictions should come true on average, for every age group, rating band and colour, and for players who have not played in a while.

This study checks that on twelve seasons of real games in two steps. First it fits small corrections on one season of results and tests them on the next, which the corrections never saw, and repeats that for every season. Then it replays the rating rules month by month, FIDE's current rules against proposed changes, and scores each the same way.

Data

Method

A season runs October to September. Every correction is fitted on one season and tested on the next, for each of the eleven test seasons from 2015-16 to 2025-26. The charts on the findings below come from the latest pair: fitted on 2024-25, tested on 2025-26. Each model predicts White's expected score before the game. Accuracy is scored with log loss and Brier score (mean squared error on 1, ½, 0). Confidence intervals come from 1,000 bootstrap resamples of whole events, because games inside one tournament are not independent.

Finding 1

FIDE's table overstates the favourite

FIDE expected score Corrected model Actual score

Finding 2

Juniors are underrated, seniors overrated

Correction in Elo points relative to players aged 21 to 49, fitted on the training year. A bar of +50 means that group plays like players rated 50 points higher.

Finding 3

Adults come back from a break weaker than their rating

Under 21 21 to 49 50 and over

A test case

Is Nakamura sitting on his rating?

Published classical rating Rated classical games in that list

FIDE standard lists, January 2024 to September 2026. A list's games were mostly played the month before it was published.

Finding 4

Ratings trail current form

Finding 5

Results

How much each correction helps

All rows are scored on the same games across the eleven test seasons, each correction fitted on the season before its test. Lower is better. The change column is the difference in log loss against the published FIDE ratings, with its 95% interval.

Every season

The pattern holds in all eleven test seasons

All corrections Age Gap scale White advantage

Reduction in Brier score against the published FIDE ratings, by test season. Each correction was fitted on the season before.

Simulation

Replaying the rules month by month

A correction that predicts well is not yet a rule FIDE can adopt. A rule changes ratings, and changed ratings feed into every later game. So each candidate rule was replayed month by month from October 2014 to September 2026, next to FIDE's current rules, using the K-factor FIDE actually applied to each player each month. The first season is a warm-up. From 2015-16 on, each season's rule uses settings fitted on the season before.

The official rating in each game already reflects every rated game in the world, and most of those are not in this sample. So each rule keeps a running difference from the official rating for every player, and that difference is what the rule changes. Under FIDE's own rules the difference stays at zero, which makes the control exact.

For FIDE

What this suggests for the rating regulations

Limits

What this study does not show

The first eleven seasons include the pandemic, when over-the-board chess nearly stopped. Those seasons have fewer games, and FIDE also made a one-off rating adjustment in March 2024.
Breaks are measured from FIDE lists that begin in September 2012, so in the earliest seasons a break of four years or more cannot be seen and is left out.
Since October 2025, FIDE's 400-point rule applies only to players rated below 2650. This study applies it to everyone, which affects a handful of elite games.
TWIC covers stronger, better-known events. The median rating here is , well above the full FIDE list, so results for club-level players need their own data.
The inactivity effects are averages over mostly non-elite players. A world-class player returning from a break may lose more or less than the average.
Birth years come from the current FIDE player list. Time since last game comes from monthly game counts, so a break is measured to the nearest month.
The simulation only updates ratings from games in this sample. A rule that FIDE adopts would act on every rated game, which the replay approximates but does not reproduce.
The junior problem is still open. Neither a larger K-factor for juniors nor an uncertainty-based K-factor helped, so juniors need a different kind of fix.

Reproduce

Run it yourself

Plain Python 3, no packages. Game files come from The Week in Chess (issues 1038 to 1664) and player data from the FIDE download page.

python3 src/fetch_twic.py     # TWIC game files, 2014 to 2026
python3 src/prepare.py        # games from TWIC, fast events removed by name
python3 src/fetch_lists.py    # FIDE monthly lists, September 2012 onward
python3 src/verify_events.py  # keeps events confirmed as classical
python3 src/backtest.py       # corrections, fitted and tested
python3 src/simulate.py       # month-by-month rule replay
python3 src/build_report.py