A Database Answers Corpus Questions
A chess database can answer questions such as “Which moves were played from this position in this selected set of games?” or “Which strong recent games reached this structure?” It does not directly answer which move is objectively best. Frequency, score and quality are different variables.
Know Which Database You Are Using
A curated master database, a huge public player corpus, your personal games and an opponent-specific database answer different questions. Their population, duplication, metadata quality and rating distribution can differ dramatically. Always interpret statistics relative to the corpus.
Sample Size Before Percentage
62% is not meaningful without a denominator. A move scoring 62% in 13 games conveys very different evidence from the same percentage in 13,000. Even large samples can be biased by rating, era, color strength, time control or repeated imported games.
Position Identity Beats Opening Names
Opening explorers are most reliable when you search the exact position rather than assuming the opening label captures move-order nuances. Transpositions mean the same position can arrive through multiple names, while similar names can conceal different positions.
Separate Frequency, Results and Recommendation
Three useful questions are:
- frequency: what do players choose?
- results: what happened in the selected games?
- quality: what does stronger analysis suggest about the moves?
A rare move can be excellent. A popular move can be outdated, practical, stylistic or simply easy to reach. Database popularity is evidence of practice, not a replacement for analysis.
Data Cleaning Is Analytical Work
Duplicates, corrupted PGNs, inconsistent names, missing dates and unreliable ratings can distort conclusions. For serious research, document filters and data source, and be cautious about precision the data cannot support.
Result Statistics Are Conditional on the Players
A move's historical score is affected by who chose it and against whom. Strong players may select one move disproportionately; an opening may be used as a surprise weapon against lower-rated opposition; a dataset may mix classical and fast games. Do not interpret raw W/D/L as if the move itself generated the outcome independently of the population.
Recent Practice and Historical Practice Answer Different Questions
Historical games can explain how ideas developed and provide excellent model positions. Recent strong games can better indicate current theoretical attention. Use both intentionally instead of applying a single date filter to every research question.
Reference Framework
A chess database is a corpus of game/position records with metadata. Different databases serve different purposes, and statistics are meaningful only relative to the actual corpus and filters.
Database percentages describe what happened in a selected sample; they do not prove what should happen with best play.
Reference / Master Database
A reference/master database emphasizes strong, curated or professionally maintained games suitable for opening/theoretical research.
Large Player Database
A large player database maximizes breadth of game records and is useful for frequency, opponent/player and practical trends with careful filtering.
Personal Database
A personal database organizes one's own games for recurring-error, repertoire and progress analysis.
Opponent Database
An opponent database collects relevant games for preparation; sample size, recency and fair-play/event rules must be respected.
Annotated Database
An annotated database adds human commentary/variations, whose author/source/date are part of the evidence.
Engine-Evaluation Dataset
An engine-evaluation dataset stores engine outputs for positions/games. It is not interchangeable with empirical game-result data.
Tablebase Is Different
Tablebase is different from all game databases because it stores exact low-material state information rather than a corpus of played games.
Duplicates
Duplicates inflate frequency, player counts and result statistics. Deduplicate or understand the source's duplicate policy before drawing conclusions.
Player-Name Normalization
Player-name normalization matters because spelling variants can split one player's games into multiple identities and corrupt statistics.
Event Normalization
Event normalization prevents the same event/location/round from fragmenting under inconsistent metadata.
Metadata Errors
Metadata errors include ratings, dates, results, event names, colors and opening tags. Statistical claims inherit those errors.
Corrupt / Incomplete PGN
Corrupt/incomplete PGN can lose moves, result, initial FEN or metadata and should be excluded or repaired with provenance rather than silently trusted.
Curated vs Raw Data
Curated data trades breadth for cleaning/selection; raw data offers scale but requires more quality control. Neither is automatically superior for every question.
Statistics Need Clean Data
Statistics need clean data because counts and filters are only as reliable as identity/result/metadata normalization. ChessBase documentation explicitly links meaningful statistics to careful editing/consistent player names.
Move Frequency
Move frequency answers how often was this move played in this filtered corpus?—not how good is the move?
White Wins / Draws / Black Wins
White-win/draw/black-win percentages are empirical outcomes from the selected games. They mix player strength, era, time control, selection and move quality.
Game Count
Game count is sample size evidence. Small samples should widen uncertainty rather than produce strong conclusions.
Average Rating
Average rating summarizes part of the player pool but can hide distribution, color imbalance and changing rating systems.
Top / Recent Games
Top/recent games are useful exemplars but are selected subsets, not replacements for the full sample.
Sample Size
Sample size matters especially when rare moves show extreme win rates from very few games.
Recency
Recency matters because opening theory and player populations change. Historical frequency and current theoretical relevance are different questions.
Rating Filter
Rating filters change the player pool and therefore the observed choices/results.
Time-Control Filter
Time-control filters matter because bullet/blitz/rapid/classical have different error rates and practical move distributions.
Player Filter
Player filters can reveal repertoire habits but create extreme selection bias if generalized to chess as a whole.
Win Rate ≠ Objective Evaluation
Win rate ≠ objective evaluation. A high-scoring move can be objectively inferior, and a theoretically strong move can score poorly in a particular sample.
Database Practice
Database practice can ask you to state database, snapshot/date, filters, sample size, move frequency/WDL and at least two reasons the statistics may not imply objective quality.