How the GOAT Score is built — the full specification

GOAT Rank scores individual athletes and teams on sporting merit, and reports a separate cultural and commercial index alongside it. The model answers a narrow question: given the public record of results, how far did this career stand above the field it actually faced? It does not attempt to answer who was most talented, most watchable, or who would win a hypothetical match across eras.

The model covers 15 sports. Comparisons inside one sport are the primary product; cross-sport comparison is an explicit opt-in with a different weighting scheme, described in section 9.

1. Published beats clever. Every constant is on this page. A model nobody can audit is an opinion with a decimal point.

2. Never invent a number. Where a statistic is unknown it stays unknown; the record is marked provisional rather than filled with an estimate that later reads as fact.

3. Separate sport from fame. Endorsements, follower counts and press volume never move the sporting ranking.

4. Same rules for everyone. No athlete is hardcoded to a score, and no result is adjusted after the fact because the output looked wrong. When an output looks wrong, the rule changes and every athlete is recomputed.

5. Versioned outputs. Each derived number carries the calculation, eligibility, normalisation and methodology version that produced it, so old numbers remain explainable.

Data enters the system through three channels, each recorded on the record itself:

Seeded editorial records. Career statistics, honours and competition results for the historically significant athletes and teams, compiled from governing-body records (FIFA, FIBA, ICC, World Athletics, World Aquatics, ATP/WTA, PGA Tour, NHL, NFL, MLB, FIA) and cross-checked against encyclopedic sources.

Structured public datasets. Wikidata claims (identity, nationality, birth date, awards, team memberships, positions), Valve's official Counter-Strike Global Standings for esports rosters, and league-published squad lists.

Weekly watch agents. Fifteen scheduled agents, one per sport, scan public sources for new athletes, teams and ranking movements and file proposals. Proposals are reviewed before publication and always land as provisional until independent confirmation of both identity and honours exists.

Rights status is tracked per source with a traffic-light gate. Only sources marked green — public, redistributable, attributable — may drive production numbers; amber and unknown sources can inform review but not scoring.

Verified. Identity confirmed and at least one documented honour or an authoritative career record. Eligible for leaderboards by default.

Provisional. The name and affiliation come from a public source, but the career record has not been independently confirmed. These records are searchable and comparable, are labelled in the interface, and score on a conservative derived scale. They are excluded from default leaderboard views.

Production eligibility for database-backed records is stricter still: the record must be published, its data quality must not be demo or estimated, and its source rights must be green.

Raw sporting statistics are not comparable, even inside a sport: a goal in 1958 is not a goal in 2024, and a batting average is not a lap time. Each sport therefore has an anchor metric and a reference cohort. An athlete's performance in a pillar is expressed as a 0–100 position against the best careers ever recorded in that same sport, not against an absolute scale.

Counting statistics are converted to rate-adjusted form where the schedule length has changed materially (games per season, races per year, tournaments per calendar), so that a longer modern season does not by itself create a higher score.

Per-sport anchors and key metrics are listed on the summary page and on each sport hub.

GOAT = 0.25·Dominance + 0.20·Peak + 0.20·Longevity + 0.20·Honours + 0.15·Statistics

Dominance (25%) — margin over contemporaries: how far ahead of the second-best of the same seasons the athlete stood. It carries the largest weight because it is the only pillar that directly measures the thing the word "greatest" means, and it is the pillar least inflated by schedule length or era.

Peak (20%) — the height of the best sustained stretch, measured over the strongest five consecutive seasons. Weighted equally with longevity so that a short, overwhelming career and a long, excellent one can both reach the top without one design choice deciding the whole ranking.

Longevity (20%) — years spent at or near the top, and the number of distinct rival generations beaten. Twenty percent inside one sport is deliberately moderate: within a single sport, peak and dominance are more diagnostic. Cross-sport the weight rises sharply (section 9), because career length is the one currency every sport shares.

Honours (20%) — titles weighted by the difficulty and field strength of the competition, not counted flat. A world title is not a domestic cup, and a domestic title in a weak league is not a domestic title in a strong one.

Statistics (15%) — normalised output: goals, points, wins, records. The lowest weight, because raw counting statistics are the most distorted by era, schedule and role, and because they are partly already reflected in dominance and peak. Giving them more weight would double-count.

There is no principled way to derive these five numbers from first principles; they are a stated editorial position, chosen to be round, stable and defensible, and held constant across all sports so that no sport gets a bespoke weighting that flatters it. The honest claim is not "these weights are correct" but "these weights are fixed, published, and applied identically to everyone".

Because the weights are a judgement call, the important question is how much they matter. Two properties hold in the current model:

First, the pillars are positively correlated — careers that dominate also tend to win honours — so moderate weight changes move scores far less than they move arguments. Second, the ranking is most sensitive at the boundary between a short overwhelming peak and a long elite career; that is precisely where the public disagreement lives, which is why the comparison tool lets you re-run any matchup over a chosen time range instead of forcing one verdict.

A formal sensitivity table — rank stability under ±5 percentage-point perturbations of each weight — is planned for version 1.1 of this paper and will be published with the raw output rather than summarised.

The depth multiplier estimates how many people realistically compete for the same place at the top. It is built from three published inputs — participation base, number of national federations, and how open the pathway from beginner to professional is — and it is applied only to cross-sport comparisons. Inside a sport it has no effect whatsoever.

Football is the reference point at ×1.00, not because it is the best sport, but because it has the largest combination of participation and worldwide professional infrastructure; every other multiplier is expressed relative to it. All participation figures are federation-published estimates, stated as orders of magnitude.

Discipline overrides. Inside some sports the pyramid differs enormously by event: effectively every child on earth sprints, while pole vault needs a pit, a coach and a pole. Where a discipline is calibrated, its multiplier replaces the sport-level value.

These multipliers are the most contestable part of the model and we treat them that way. They are estimates of participation, not verdicts on quality, and they are the first thing we expect to revise when better participation data or an outside reviewer's argument arrives.

Cross-sport comparison uses a different pillar weighting from the in-sport model, because the pillars are not equally comparable between sports. Honours counts are sport-specific artefacts; sustained excellence is not.

The depth factor is then applied as a partial multiplier. Only part of the score is scaled by the talent pyramid; the majority reflects performance regardless of sport, so a golfer is never mechanically capped below a footballer.

Cross-sport GOAT = Σ(pillar × cross-sport weight) × (0.65 + 0.35 × depth^1.25)

With the exponent at 1.25, the full range of depth multipliers (×0.76 to ×1.00) produces a spread of roughly 10.2% between the shallowest and deepest sport — enough to matter at the very top of a mixed list, not enough to decide it on its own.

The summary methodology page states this as "in-sport score × competition depth". That is a simplification of the formula above; this paper is the normative version.

Era adjustment is applied by measuring against contemporaries rather than by applying a global correction factor to old careers. Dominance and peak are computed against the field an athlete actually faced, which handles most of the problem: a 1960s athlete is compared with 1960s rivals, not with 2020s equipment and sports science.

Schedule inflation is handled in normalisation (section 5): longer modern seasons are rate-adjusted so accumulating more games does not by itself raise a score.

In the separate cultural index, pre-social-media careers receive an era projection of contemporary media reach instead of a follower count, so that Ali, Pelé or Bradman are not penalised for being born before Instagram.

What the model deliberately does not do is apply a blanket "standard of play has risen" discount to historical athletes. That adjustment cannot be measured, only asserted, and it would silently encode a preference for the present.

Any comparison can be re-run over a restricted window: all-time, best five consecutive years, a calendar decade, the overlapping prime era of the athletes compared, or a custom range. Scores are recomputed, not filtered.

Within a window, each pillar is rescaled by how much of the career the window captures:

intensity = window average ÷ career average · share = window output ÷ career output · span = window seasons ÷ career seasons

Dominance scales with intensity, longevity with span, and honours and statistics with share, each within clamped bounds so a single freak season cannot produce an impossible score. The peak window itself is found by scanning for the best five consecutive seasons by summed output.

Careers start in different calendar years, so the pace tools run on a career clock rather than a calendar: Career Day is the number of days since that athlete's own first professional appearance. Day 100 of a 1997 rookie is compared with day 100 of a 2026 rookie.

Performance Share Index (PSI) expresses a result as the share of the available field taken — for example prize money won divided by the total purse. Missed cuts, withdrawals, disqualifications and non-starts are recorded as zero, not as missing. Where currencies or fields do not match, PSI returns nothing and states why; it is never approximated.

Career Greatness Index (CGI) is a percentile-weighted composite: each component is scored against the historical pool measured at the same career day. When a pool is unavailable the remaining weights renormalise, so no athlete is ever hardcoded to 100.

Ghost Race reports cumulative output at fixed career-day checkpoints (30, 100, 365, 1 000, 1 825 and 3 650 days), which is what makes "who was further ahead at the same stage" a measurable question rather than a memory test.

These pace tools inform the career timeline and comparison views. They are reported alongside the GOAT Score; they do not silently feed into it.

Fame is measured, published, and kept out of the sporting ranking. The cultural index is a composite of five components:

Total legacy = Cross-sport GOAT × 82% + Cultural & Brand Impact × 18%

Legacy is an optional view. The default ranking everywhere on the site is pure sporting data, with the cultural index reported as its own column so the two can be read against each other instead of being blended out of sight.

Scores are reported to one decimal. That precision is a display convention, not a claim of accuracy: differences below roughly one point should be read as a tie, and the interface treats near-equal scores as a contested verdict rather than a winner.

Confidence varies by record. A seeded athlete with complete competition results and verified honours carries far more weight of evidence than a provisional record built from a squad list. Both are labelled, and only the former appears in default rankings.

Coverage is uneven by sport and by era. Football, tennis, basketball, golf and athletics have the deepest verified records; volleyball, table tennis and parts of esports have the thinnest. Uneven coverage affects who can be compared reliably, not how the formula treats them.

Active careers. Longevity is scored on career to date. An active athlete's score can rise; it is a snapshot, marked as active.

Interrupted careers. Seasons lost to war, bans, boycotted Olympics or catastrophic injury are recorded as missing, not as zero. Missing seasons reduce accumulation but do not count against per-season quality.

Team sports and individual credit. Team honours are weighted by competition strength and by the athlete's role, and never treated as identical to individual titles.

Golf and other cut-based sports. Events where a player misses the cut score zero share rather than being excluded, so consistency of qualification is reflected rather than hidden.

Motorsport. Results depend heavily on machinery. Dominance is therefore weighted towards intra-team and relative-field measures, and the depth multiplier already reflects the closed pathway.

Esports. Titles change, games are patched, and careers are short. Longevity is scored against the realistic lifespan norms of the discipline rather than against football-length careers.

Doping and sanctions. Officially stripped results are removed. Allegations that never produced a sanction are not scored, in either direction; the sources are linked and readers can draw their own conclusion.

Men's and women's competition. Scored within their own competitive fields, with a gender filter available. Mixed lists compare positions within field, never raw performance across fields.

1. The weights are editorial. They are round numbers chosen for defensibility, not derived from data. A reasonable person could argue for 30/20/20/20/10 and get a different order at the margin.

2. Depth multipliers are estimates. Participation figures are self-reported by federations, inconsistently defined and often years out of date. This is the weakest input in the model.

3. Modern-era data bias. Recent careers have richer, more granular records. Better data can look like better performance.

4. Anglophone source bias. Encyclopedic coverage over-represents athletes written about in English, which affects discovery and verification speed, particularly outside Europe and North America.

5. Team-sport attribution. Separating an individual's contribution from their team's quality is an unsolved problem in every public model, including this one.

6. Cross-sport comparison is inherently arguable. We publish it because people want it, with a different weighting and an explicit opt-in, not because we think it is as sound as an in-sport comparison.

7. No formal sensitivity analysis yet. Planned for v1.1, as noted in section 7.

Every derived number carries version stamps so an old score stays explainable. Current stamps: calculation calc-v1, eligibility elig-v1, normalisation norm-v1, methodology method-v1.

Weights and multipliers change only with a version bump and a dated note in the change log below. Scores are never adjusted for individual athletes outside a rule change.

We are actively looking for statisticians, sports researchers and federation analysts to review this specification — particularly the competition-depth derivation in section 8 and the sensitivity work planned for section 7. Reviewers are credited by name in the paper unless they prefer otherwise.

Factual corrections to any athlete or team record are equally welcome and are treated as higher priority than feature work. Include the source and we will correct it, mark the record, and note it in the change log.

Write to hello@goatrank.world or use the contact page.

GOAT Rank (2026). GOAT Rank Methodology Paper, Version 1.0. goatrank.world/methodology/paper

This paper may be quoted and criticised freely with attribution. If you publish a rebuttal, send it to us — we will link it from this page.