Goal Constellation

Goal Constellation is the first module of the World Cup Memory Machine, a multi-piece data project on the history of men's FIFA World Cups (1930–2022). This methodology section documents every source, transformation and acknowledged limitation. Transparency is a feature, not a footnote.

Data sources

Tier A — exact coordinates (390 goals, 88.4%): sourced from StatsBomb Open Data, competition_id 43 (FIFA World Cup), seasons currently published: 1958, 1962, 1970, 1974, 1986, 1990, 2018, 2022. Includes (x, y) location in StatsBomb's 120×80 pitch reference frame, body part, shot type, technique and StatsBomb xG model output where computed.

Tier B — reconstructed zone (51 goals, 11.6%): manually curated from FIFA archival reports, contemporary press footage and Wikipedia structured tables for editions not yet in open data: 1930, 1934, 1938, 1950, 1954, 1966, 1978, 1982, 1994, 1998, 2002, 2006, 2010, 2014. Coordinates approximate the zone where the goal was scored (six-yard box, penalty area, edge of area, midfield, etc.), not the precise pixel. Each entry carries a narrative note traceable to public footage.

Shots (1,958 misses): StatsBomb event data from 2018 + 2022 only (the two tournaments with complete shot coverage). Filtered to outcomes Saved, Saved Off Target, Saved to Post, Post, Off T, Wayward. Blocked shots excluded — too common, less narratively informative.

Coordinate transformation

StatsBomb publishes (x, y) where x ∈ [0, 120] is the length axis (x=120 is opponent goal line) and y ∈ [0, 80] is the width axis. The constellation uses a normalized pitch in percent: x ∈ [0, 100] lateral, py ∈ [0, 100] longitudinal, with py=100 at the attacking goal. Transform: x' = sb_y / 80 × 100, py' = sb_x / 120 × 100. All teams are unified to attack the top goal regardless of original home/away orientation.

Filtering decisions

• Penalty shootouts excluded. Period ≥ 5 events removed (50 shootout goals across 2018 + 2022). These all happen at one spot and would create a meaningless column.
• Penalty kicks (open play) jittered. Penalties land at (50, 88) by definition. To avoid visual collapse, each is offset by ±2.5 in (x, py) using a deterministic hash of scorer + year + minute. Same penalty always renders at the same pixel; never random.
• Names normalized. Full registered names compressed for display (e.g. Edson Arantes do Nascimento → Pelé, Eduardo Gonçalves de Andrade → Tostão, Jair Ventura Filho → Jairzinho). Full names preserved in the underlying dataset for traceability.
• Country names abbreviated. Korea Republic → S. Korea, West Germany → W. Germany, United States → USA.

Layers and visual encoding

Tier A goals: solid white/cool/warm dots according to phase (group/knockout/final), opacity 0.85. Tier B goals: warm-orange translucent dots with diffused halo, opacity 0.55. Posts: red empty rings. Saves: dim blue-grey filled circles, opacity 0.28. Off-target: even dimmer dark grey, opacity 0.18. xG mode overrides phase colouring with a purple-to-warm gradient based on each goal's expected probability.

The xG model

xG (expected goals) values are from StatsBomb's proprietary model: probability a shot becomes a goal given its location, context, body part, defensive pressure and other factors. Values range 0.001 (impossible) to 0.99 (sitter). 325 of 441 goals in this dataset carry xG — only those from 2018 and 2022 events with explicit statsbomb_xg field. The "Wonder goals" preset filters xG < 0.05 (less than 1-in-20 chance).

Acknowledged limitations

• Tier B coverage is sparse: 1930–54 has only 2–3 reconstructed goals per tournament. Editions like 1994 USA and 2010 South Africa lean on a handful of iconic moments — they do not represent every goal.
• The 1990 StatsBomb data is a single match — likely the final West Germany vs Argentina. Other 1990 goals are absent from this prototype.
• xG model has improved over time; 2018 values are computed by an older model than 2022.
• Misses are only available for 2018 + 2022. Older tournaments show no shot layers when activated.
• "Conversion rate" metric is computed only over the goals + misses currently visible. It is sample-dependent, not a tournament average.

Technical stack

Single-file HTML/CSS/JS. No build step, no framework. Canvas 2D for goal rendering with device pixel ratio for retina sharpness. SVG for the pitch geometry (sharp lines at any scale). Web Audio API for goal tones (sine waves, three pitches keyed to phase). URL state encoded in query string; current view fully reproducible from URL alone. ~308 KB total payload including 441 goals + 1,958 misses inline.