Openings › How it works
How the numbers are made
- Source: Lichess open database, the 12 monthly files from Oct 2025–Sep 2026, released under CC0 1.0. Chess Opening Scores is independent and not affiliated with or endorsed by Lichess.
- Sample: 90 slices spread evenly through each month's file, pooled over the 12 newest monthly files (a rolling 12-month window: each new month replaces the oldest), covering games from 1 Oct 2025 – 30 Sep 2026 (362 distinct days). The same reproducible sample from each of the 12 newest monthly files, pooled: 90 slices per file at fixed byte offsets spread evenly across it, offset_k = (k + 0.5) / slices * file_size, each decoded from its first complete compressed block. Same months + same settings = same games. Games used by month: Oct 2025 412,960; Nov 2025 490,274; Dec 2025 512,080; Jan 2026 467,359; Feb 2026 482,874; Mar 2026 447,260; Apr 2026 491,683; May 2026 435,429; Jun 2026 490,248; Jul 2026 462,614; Aug 2026 480,661; Sep 2026 456,087. Every month of the window was sampled with 90 slices.
- Games read vs. used: 6,021,814 games read; 5,629,529 used after filters (rated bullet, blitz, rapid, classical games; players within 200 rating points; excluded terminations: Abandoned, Rules infraction). Excluded: rating 296,285, speed 64,467, termination 31,520, result 13, unreadable moves 0. Games whose first moves cannot be read are counted as read but excluded from "used", from rating-level totals and from every percentage.
- Time controls: as labelled by Lichess (bullet, blitz, rapid, classical). Rating-level and time-control tables are separate summaries; neither is split by the other.
- Lichess ratings: ratings come from Lichess and are not interchangeable with Chess.com or FIDE ratings.
- Data period vs. page updated: the data period is when the sampled games were played (1 Oct 2025 – 30 Sep 2026). "Page updated" is the date of the last meaningful data or content change (19 Oct 2026).
- Reproducibility: files https://database.lichess.org/standard/lichess_db_standard_rated_YYYY-MM.pgn.zst for 2025-10, 2025-11, 2025-12, 2026-01, 2026-02, 2026-03, 2026-04, 2026-05, 2026-06, 2026-07, 2026-08, 2026-09, each identified by its size and ETag (recorded with the monthly summaries in our repository), 90 slices of 4,194,304 bytes per file; offset_k = int((k + 0.5) / slices * file_size).
- Rating level: average rating of both players. Levels have no gaps: each covers from its lower limit up to, but not including, the next level's lower limit (e.g., an average of 1199.5 is "Under 1200").
- Matching openings: a game counts for an opening if it reached the same position after the same number of moves — including different move orders. Positions reached after a different number of moves are not counted.
- Minimum sizes: 300 games to show a rating level, 150 games to show a reply. A page is published only with at least 3 rating levels shown and 2,000 games in total.
- Margins of error (±): approximate 95% margins based only on the number of games. They do not cover sampling bias (games come from specific moments of each month) or related games (the same players play many games), so real uncertainty is larger. Small differences may be chance, and these comparisons alone can't show that one reply is reliably better.
- What the numbers mean: observed results for the selected games. They don't prove that an opening causes wins — stronger or weaker players may simply prefer certain openings.
- Text: every sentence on the opening pages is generated by code from these numbers. The wiki pages and the wiki sections of opening pages are the exception: their explanations come from named sources (see How the wiki texts are made).
- Monthly reports: once a month, a report compares the newest month's games with the month before (two separate monthly samples, never two overlapping windows). It lists only changes of 3 standard errors or more that are also large enough; a quiet month says so. Each report is frozen when it is first published. Newest: the latest monthly report (Sep 2026).
- Downloads (CSV, CC0): each opening page links a file with the same numbers as its tables. Columns: summary (rating level, time control or reply), level or move, games, White win / draw / Black win %, score %, approximate ±95% margin, and whose score it is ("score_for"). Groups with too few games keep their count but have empty percentages, as on the page.
How the mate-in-one exercises are chosen
- Source: positions from rated games in our sample of the Lichess open database (CC0 1.0), the same stored monthly samples as the opening pages. We download nothing extra for them.
- Games: we leave out the terminations and the speeds our other pages leave out (UltraBullet and correspondence games are not used); the other rules our statistics use to leave out games are not applied here.
- From each stored slice of a monthly Lichess file we take at most one position, and the slices are spread over the month's file.
- The side to move has at least eight legal moves, and at most three of them give checkmate.
- The game's mate came at its 20th half-move or later, which leaves out quick opening traps.
- We leave out a bare king against nine or more points of material (pawn 1, knight 3, bishop 3, rook 5, queen 9): the end of a won ending.
- Within a slice we look at the first 8 positions that pass these rules and keep the one whose side to move, then whose mating piece, has the fewest exercises so far, so that both colours and different pieces appear.
- Duplicates: a position that is already an exercise is never chosen again, also in a later month.
- Checks at every build: the position is valid and the game not over; the side to move is right; every legal move is tried and the list of mating moves must be complete; the last move shown must be legal and lead to the position; the hints and the explanations are computed from the position again by the site checks.
- At most 50 new exercises per processed month. Exercises are never removed, and their addresses never change.
- The pages make no claim about difficulty, popularity, ratings or the players, and show no player names and no links to the games. Everything on them is assembled by code.
The exercises: Mate in one.
How the weekly chess roundup is made
Source: Event facts and opening statistics counted by ChessOpeningScores from Lichess official broadcast games (lichess.org; full exports: https://database.lichess.org/#broadcasts), licensed CC BY-SA 4.0: https://creativecommons.org/licenses/by-sa/4.0/. Changes: we counted games dated in this week and computed opening statistics; player names, ratings, titles, federations, FIDE IDs, clocks, engine evaluations and comments removed. These are counts of recorded broadcast games, not official results or standings. This section is shared under CC BY-SA 4.0, with no warranty. Not affiliated with or endorsed by Lichess.
Counting rules
- Source: the game files of official Lichess broadcasts, read once a day; only counts are kept, never a game or a player's name.
- Counted games: standard chess with a result (1-0, 0-1 or ½-½), dated inside the week (Monday to Sunday, UTC), without a BOT player; repeated games inside a round are counted once.
- An event is all rounds of one Lichess broadcast in the week; broadcasts whose names share the text before " | " are one event with sections.
- An event is held back, without its name, when its name fails the name check (a player's name from its games, a youth or age-group event, words the site's wording rules exclude, or no name) or when a finished round of the week is missing.
- An event is shown with at least 10 counted games; an issue shows the 10 events with the most counted games and needs at least 5, so a quieter week gets no page.
- Every number is a count of recorded broadcast games. Corrections stay visible on the page, dated, with the reason.
Formulas
- Counted games (N): standard chess games with a result (1-0, 0-1 or ½-½), dated inside the week (UTC), without a BOT player; repeated games inside a round are counted once; unfinished games are not counted.
- Share = games with that result ÷ N, as a whole-number percentage; each share is rounded on its own, so the three may not add up to 100.
- Decisive games = (White wins + Black wins) ÷ N.
- Average length = total half-moves ÷ 2 ÷ N, in full moves, rounded to a whole number.
- Time control (Lichess's formula): estimated seconds = base time + 40 × increment of the first period; under 180 s bullet, under 480 s blitz, under 1,500 s rapid, otherwise classical; a time control that cannot be read is unclassified. Organisers often write the base time in minutes, so a base of 120 or less (for example 90+30) is read as minutes and a larger one (for example 5400+30) as seconds. An event shows a class when at least 90 % of its classified games share it, otherwise "mixed", and "not classified" when unreadable time controls are the largest group.
- Most played opening: the family (the name before ":") of the deepest named position each game reached, from the site's CC0 opening table; shown with at least 5 games and only without a tie.
- Opening families table: families with at least 20 games across the events of the issue, at most 8, by games.
- First moves: White's first move in each counted game; share = games with that move ÷ games with a first move.
Limits
- Only official Lichess broadcasts are read; other events and other sites are not covered.
- The game files come from the organisers' feeds and can be corrected later; a correction found later is shown on the issue with its date.
- These are counts of recorded broadcast games, not official results.
- A round that had not finished at least 6 hours before an issue was made is left out of that issue, and the number of such rounds is shown.
- Opening families come from the site's opening table; a game counts under the deepest named position it reached.
Corrections
An issue is frozen when it is first published and never computed again. A correction keeps the event on the page, struck through, with the date and the reason, and the issue lists it under Corrections; nothing is changed silently.
The issues: Weekly chess roundup.
How the wiki texts are made
- Every sentence that explains chess on a wiki page or in a wiki section carries a numbered source. The source list gives each source's address, its publisher, the day we read it and its licence.
- We read sources to check facts and write our own wording; we do not copy their text. A check compares every sentence with stored fingerprints of the sources' phrases and stops the build if eight words in a row match.
- A strategic idea needs two sources from different publishers, one of them a teaching, specialist or primary source. History needs a specialist or primary source. When sources disagree, the claim is held back.
- Moves and positions are replayed by code (python-chess). Every line shown is a named line of the free opening list or was played in at least 150 games of our sample.
- Statistics come from the same 12-month sample of rated Lichess games as the opening pages, calculated by code with the same minimum game counts; dates and history come from the cited sources. A sentence that depends on a number is left out automatically when the number no longer supports it.
- The explanations were drafted with the help of an AI model and then checked against these rules. Separate AI review sessions, which may use the same AI model, check every sentence against its sources; a sentence they cannot support is held back. No human chess expert has reviewed the texts yet.
- Errors can be reported from every page.
The wiki: Chess wiki. How we count move orders: Transposition explained.