Methodology 0.3.0-research-preview
Nothing hidden.
A configurable 100-point framework aligned to the post-2022 Ballon d’Or criteria. The engine computes; the AI explains.
Individual performance
55
Position-specific output, creation, all-round contribution, execution, availability, and consistency from real match data.
Decisive moments
25
Performance against strong pre-match Elo opponents, in knockout rounds, and in verified score-changing situations.
Team achievements
20
Verified finishes allocated by competition prestige, route, participation, and within-team importance.
Component model
Nine named contributions.
Within-position percentiles make forwards, midfielders, defenders, and goalkeepers comparable without pretending their jobs are the same.
Role output
signal: role_output_signal
Creation & progression
signal: creation_progression_signal
All-round role contribution
signal: all_round_role_signal
Efficiency & execution
signal: execution_signal
Availability & consistency
signal: availability_signal
Strong-opposition performance
signal: strong_opposition_signal
Knockout & final impact
signal: knockout_finals_signal
Match-swinging contribution
signal: match_swing_signal
Team performance & achievements
signal: team_achievement_signal
Vote forecast
Model the ballot without calling popularity merit.
The main Index remains a 100-point football score. A separate statistical forecast will estimate how the journalist electorate may vote after regime-aware calibration.
Status: held · no unvalidated blend is published
Collective & tournament narrative
3 / 10A calibrated vote-model feature for decisive trophies and tournament prominence. It remains separate from the normative merit score.
Empirical position prior
2 / 10A regime-aware coefficient learned from historical voting, never applied to the position-neutral football merit score.
Club & league visibility
1 / 10Season-specific ballot visibility, distinct from team quality and competition strength.
Recognition & aura
1.5 / 10Time-decayed prior Ballon d'Or voting recognition; no manual reputation score, follower count, or subjective vibes.
Media momentum
2.5 / 10Source-weighted candidate prominence in audited football journalism during the eligibility window, excluding paid promotion and raw social volume.
Forecast coefficients must be fitted on historical ballots and improve out-of-sample rank agreement without creating unacceptable club, league, nationality, or position bias. Prediction markets are a benchmark, not a hidden scoring input. Missing evidence blocks the forecast; the LLM never fills it.
Position adjustment
Compare role to role first.
Forward
Role output: non penalty goals per90 35% · non penalty goals 20% · assists per90 15% · assists 15% · shots on per90 15%
Midfielder
Role output: assists per90 20% · assists 15% · key passes per90 25% · non penalty goals per90 10% · accurate passes per90 20% · shots on per90 10%
Defender
Role output: tackles per90 20% · interceptions per90 25% · blocks per90 15% · duels won per90 20% · accurate passes per90 10% · goal contributions per90 10%
Goalkeeper
Role output: save rate 30% · saves per90 15% · clean sheet rate 20% · goals conceded inverse 20% · penalties saved 15%
Trophy multipliers
Achievement is weighted, never binary.
“Team” means achievement plus the player’s contribution to that achievement: competition stage and prestige, route difficulty, participation, and importance. A squad medal alone cannot receive full credit. Club and league visibility are excluded here and live only in the separate vote forecast.
Validation gate
Modern-regime calibration.
Hold out 2024–2025.
The target is Spearman ρ ≥ 0.70 against the official Ballon d’Or top ten, plus top-five Jaccard and market-baseline comparison. The holdout years are not used to tune weights.
Season-based editions from 2022 onward are the primary calibration regime. Earlier editions inform historical priors only through explicit regime flags; raw vote totals are never compared across voting-system changes.
Every rank divergence is published with its component-level explanation. A disagreement is a football argument to inspect, not an error to hide.
No historical race is marked validated until the required API coverage and official reference rankings pass audit.
Missing data
Graceful degradation.
An unavailable football signal receives no imputed credit, lowers the player’s data-completeness score, and is identified in the component sentence. Scores with material gaps remain non-publishable. Missing jury context blocks the vote forecast entirely.
AI boundary
The model explains. It never scores.
Build-time explanations receive the final rank, named component decomposition, and only timestamped shortlist, market, or reporting evidence that passed the source schema. Vercel AI Gateway selects the model by configuration. The deterministic engine remains provider-independent and the model has no scoring authority.