{ "fetched_at": "2026-10-20T02:25:40Z", "source_url": "https://www.kaggle.com/competitions/chess/data", "pages": [ { "name": "Custom", "content": "When predicting the outcome of chess games, you typically need two things; a rating system wherein the current ability of each player is estimated based on past results, and a model for estimating the expected score for each player, once you know their ratings.

Most rating systems use some methodology to determine initial \"seed\" ratings for the pool of players, and then update those ratings based on ongoing results.  The most famous approach is the Elo approach, where the applied change to a player's rating is proportional to the amount by which they exceed their aggregate expected score across all their recent games.  The scaling factor is known as the \"K-factor\", and for the official ratings used throughout the world, the K-factor is highest for new players and lowest for topmost players.  But there are many other approaches: the Ken Thompson approach takes each player's most recent 100 games and calculates the rating that would be most likely to lead to that performance.  The Mark Glickman approach is similar to Elo but introduces additional parameters for each player, tracking the level of confidence and level of volatility for each player's rating, and then using these parameters to determine which K-factor to apply.

The initial seed ratings are typically determined through a simultaneous calculation: a start rating is assumed for each player, then a \"performance rating\" is calculated for each player based on their results and the ratings of their opponents, and then those performance ratings are fed back into another iteration as the start ratings.  This is allowed to run until it converges upon a stable set of ratings.  This was the methodology used to calculate initial ratings for most major rating systems.  In fact this is the overall approach taken by the Jeff Sonas Chessmetrics rating calculation, and is used not just to calculate initial ratings but in fact to calculate all ratings.

There is a general convention in chess rating systems whereby the difference in two ratings is used for calculating expected score when two players face each other.  Of course it could just as well be the ratio of the two ratings, or some other more complex relationship that depends on the magnitude of the ratings and not just their relative difference.

Here are some links to articles existing on rating systems:
Elo and Ken Thompson
Glicko
Chessmetrics
Microsoft TrueSkill

Jeff Moser has a C# implementation of Elo and TrueSkill on Github.
Jason Brownlee has posted a Java implementation of Glicko on Github.
", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false }, { "name": "Custom", "content": "This competition was organized by Jeff Sonas of Chessmetrics.
", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false }, { "name": "data-description", "content": "The dataset of chess results represents 105 months' worth of actual game-by-game results among 8,631 of the world's top 13,000 chess players, from sometime in the last 12 years. Out of the 8,631 players included in the dataset, only 70% of the games among those players have been included. Therefore, these players actually play with approximately twice the frequency that you see in this dataset.

The players are uniquely identified by ID #’s ranging from 1 to 8,631.  The dataset is divided into a training dataset, representing a consecutive stretch of 100 months of game-by-game results among those top players, and a test dataset, representing the next 5 months of games played among those players (obviously the actual game-by-game results on the test dataset have been withheld). 

You should use training_dataset.csv to train your models. It includes 65,053 rows of data, representing 65,053 distinct games played from months 1 through 100, with the following columns:

Month # (from 1 to 100)

White Player # (from 1 to 8,631)

Black Player # (from 1 to 8,631)

Score (either 0, 0.5, or 1)

“White Player” represents the ID # of the player who had the white pieces, and “Black Player” represents the ID # of the player who had the black pieces.  The possible values for Score represent the three possible outcomes of a chess game (1=White wins, 0.5=draw, 0=Black wins). 

In chess, the player with the white pieces gets to move first and therefore has a slight advantage.  For instance, in the 65,053 games listed in the training dataset, White won 32.5% of the games, Black won 23.4% of the games, and 44.1% of the games were drawn (draws are very common among top players)

The test_dataset.csv should be used to frame submissions. It includes 7,809 rows of data, representing 7,809 distinct games played from months 101 through 105, with the following columns:

Month # (from 101 to 105)

White Player # (from 1 to 8,631)

Black Player # (from 1 to 8,631)

Score (either 0, 0.5, or 1)

The format of this dataset is the same as training_dataset.csv, except that the results for the Score column are not provided.  Competitors should calculate the white player's expected score (between 0 and 1) for each row. Once they have filled in the score column for test_dataset.csv, then can enter their submission. example_submission.csv gives an example entry.

Note: the public leaderboard is calculated based on 781 matches in the test dataset (these 781 matches are not used in the calculation of the final standings). The rest of the test dataset is used in the calculation of the final standings.

To enter your submission, visit the Make a Submission page.

Update: Several participants have noticed that the five-month test set and the final five months of the training set have some different characteristics.  This is because some filtering was done to the test set.  The filtering was done for good reasons (see the forum topic Cross Validation Dataset for more details) but it can make cross-validation challenging.  Therefore we have decided to release a cross validation dataset (cross_validation_data.csv - downloadable below), a subset of the month 96-100 training games that should more closely resemble the characteristics of the test set.  The cross validation dataset has been created, from months 96-100 of the training dataset, by excluding any games where one or both players had played fewer than 12 \"fully-rated\" games across months 48-95 of the training dataset, where a \"fully-rated game\" is one where both players already had a FIDE rating at the time the game was played.  We expect that this file can be productively used for cross validation where months 96-100 are treated like the test set.
", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false }, { "name": "Description", "content": "The Elo rating system was invented half a century ago by Hungarian-born physicist and chess master Arpad Elo. It is the most famous technique for rating chess players and is used throughout the chess world. It has been applied to many other contests as well, including other board games, sports, and video games.  However, it has never really been demonstrated that the Elo approach to calculating chess ratings is superior.  Elo's formula was derived theoretically, in an era without large amounts of historical data or significant computing power.  With the benefit of powerful computers and large game databases, we can easily investigate approaches that might do better than Elo at predicting chess results.

There are several alternatives to the Elo approach. Professor Mark Glickman developed the Glicko and Glicko-2 systems, which extend the Elo system by introducing additional parameters to represent the reliability and volatility of player ratings.  Ken Thompson uses a linearly weighted average of a player's last 100 results to calculate a weighted performance rating.  Jeff Sonas (who put together this competition) developed Chessmetrics ratings to maximize predictive power. More details are available on the hints page.

We want to see if somebody out there can do even better.  Competitors train their rating systems using a training dataset of over 65,000 recent results for 8,631 top players. Participants then use their method to predict the outcome of a further 7,809 games.

Along with the opportunity to help shape the future of chess ratings, the top ten entries win the following prizes (assuming they share their methodology).
1. Fritz DVD autographed by world champions Viswanathan Anand, Garry Kasparov, Anatoly Karpov and Viktor Korchnoi (see image)
2. ChessBase 10 Starter Package
3. Big Database 2010
4. Kasparov: How to play the Queen's Gambit
5. Anand - My Career (Vol 1 and 2)
6. Kramnik - My Path to the Top
7-10. $50 Amazon.com voucher
Prizes 1-6 have been donated by Chessbase.

Update: The team Elo Benchmark (see the leaderboard), uses the Elo rating system.

Update 2: Thanks to Chessbase, we've got some new prizes up for grabs.

Update 18 October: Some of the rules of this competition have not been explcitly stated, notably that:
1. participants are restricted to two entries per day;
2. the public leaderboard is calculated based on 20 per cent of the test dataset. The final standings are calculated based on the remaining 80 per cent of the test dataset; and
3. the final standings are calculated based on participants' best (rather than last) entry. The best entry is the entry that scores highest on the 80 per cent of the test dataset that is used to calculate the final standings (which may different from the best entry on the public leaderboard).
", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false }, { "name": "Evaluation", "content": "Entries are scored based on how accurately the entrant manages to predict the score per chess player per month.

Entrants make predictions on individual games, and those predictions are then aggregated on a by chess player by month basis. An entry's score is the RMSE on chess players' expected monthly score.

The scoring method is shown using example below.

Table 1 - a sample by game dataset with three chess players, #1, #2 and #3:
\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\r\n\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\r\n\r\n
Month #White Player #Black Player #Predicted ScoreActual Result
101120.181
101130.351
102210.480
103120.290.5
104210.230.5
105120.271
\r\n
The calculation is made in several steps.

Step 1 - chess players' predicted scores and actual scores are summed by player by month. Notice that player #1 played two games in month 101 (see Table 1). So player #1 in month 101 (row 1 of Table 2) has a predicted score that is the sum of their predicted scores from both games (they were given a probability of 0.18 and 0.35 of winning each game, so the sum of their predicted scores is 0.53). Player #1 in month 101 (again row 1 of Table 2) has an actual score that is the sum of their actual scores from Table 1 (they won both games so the sum of their actual scores is 2).

Table 2 - predicted and actual scores by player by month
\r\n\r\n\r\n\r\n\t\r\n\t\r\n\t\r\n\t\r\n\t\r\n\t\t\r\n\t\r\n\t\r\n\r\n\r\n\t\r\n\t\r\n\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\t\r\n\t\t\r\n\t\r\n
Month #PlayerPredicted ScoreActual ScoreSquared Error
10110.5322.16
10120.8200.67
10130.6500.42
10210.5210.23
10220.4800.23
10310.290.50.04
10320.710.50.04
10410.770.50.07
10420.230.50.07
10510.2710.53
10520.7300.53



RMSE0.68
\r\n
Step 2 - the squared error (in column 5 of Table 2) is calculated as: (actual score - predicted score)^2.

Step 3 - the root mean squared error (at the bottom of Table 2) is calculated as the square root of the the average squared error.

Note: the public leaderboard is calculated based on 20 per cent of the test dataset.
", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false }, { "name": "Rules", "content": "\r\n\t\r\n

One account per participant

\r\n

You cannot sign up to Kaggle from multiple accounts and therefore you cannot submit from multiple accounts.

\r\n

No private sharing outside teams

\r\n

\r\n Privately sharing code or data outside of teams is not permitted.\r\n It's okay to share code if made available to all participants on the forums.\r\n

\r\n

Team Mergers

\r\n

\r\n Team mergers are allowed and can be performed by the team leader. In order to merge, the combined team must have a total submission count less than or equal to the maximum allowed as of the merge date. The maximum allowed is the number of submissions per day multiplied by the number of days the competition has been running.
\r\n

\r\n\r\n

Team Limits

\r\n

There is no maximum team size.

\r\n

Submission Limits

\r\n

You may submit a maximum of 2 entries per day.

\r\n

You may select up to 2 final submissions for judging.

\r\n

Competition Timeline

\r\n

Start Date: 8/3/2010 12:00 AM UTC

\r\n

Merger Deadline: None

\r\n

Entry Deadline: None

\r\n

End Date: 11/17/2010 8:00 PM UTC

\r\n\t\r\n", "isPublished": false, "hasIsPublished": false, "postTitle": "", "hasPostTitle": false, "mimeType": "", "hasMimeType": false } ] }