Deep insights
Model explainability and dataset analytics, computed live over the full World Cup 2026 database.
150 years of international football
Era-level trends computed live over 49,256 international matches (1872–2026) and the full goalscorer record — scoring inflation, home advantage, goal timing and the most common scorelines in the sport’s history.
The defensive era: average goals per match has roughly halved since football’s wild early decades.
Home-win % (non-neutral venues) has hovered near 50% for a century; the draw rate is the quiet constant.
Across every recorded goal — teams score most in the final 15 minutes, when legs tire and games open up.
The tight 1–0 edges out the 1–1 and 0–0 — international football is a low-scoring game.
- 1–05,083
- 1–14,891
- 0–03,957
- 2–03,825
- 2–13,760
- 0–13,443
- 1–22,542
- 3–02,352
How the model decides — feature attribution
Each point is one of 72 fully-resolved fixtures; position is that feature’s signed contribution to the home-win logit, colour is the feature value (low → high). Bars show global mean |contribution|.
How a prediction is built — the causal chain
The full pipeline from raw data to outcome probabilities. Edge thickness is each feature’s real mean |contribution|; colour shows whether it tilts the result toward the home or away side.
Signal correlation — what overlaps
Pearson correlation between the model’s team-level inputs across all 48 nations. High values mean two signals carry overlapping information (multicollinearity) — useful for understanding what really drives a prediction.
Predicted outcome mix
Across 72 matches. The draw share is structurally inflated — draw has a fixed logit of 0 while both win logits start at −0.35.
Win probability vs ELO edge
P(home win) against ELO difference; point size encodes prediction confidence, colour the predicted outcome.
Squad market value vs ELO rating
All 48 qualified nations; ordinary-least-squares fit on log₁₀(value). Colour by confederation.
The Title Race, Visualized
Every contender as a bubble — ELO strength (x), model title probability (y), squad market value (size), confederation (colour). Spot favourites, value outliers and regional clusters at a glance.
Contenders across every axis
Each line is a team across the model's key metrics (title odds, ELO, form, win rate, squad value, World Cup pedigree). Hover to trace one team and spot trade-offs a single chart hides.
The field by confederation
Inner ring = confederations sized by qualified teams; outer ring = the teams, ordered by ELO. Hover any wedge for detail.
Top-4 strength profiles
Five dimensions normalised 0–100 across all qualified nations.
ELO power ranking (top 16)
Radial bars sized by ELO rating, coloured by confederation.
Geographic distribution of qualified nations
All 48 nations at their real coordinates, sized by ELO and coloured by confederation — the global spread of strength heading into the tournament.
ELO distribution by confederation
Box = inter-quartile range with median; whiskers = min/max; dots = individual nations.
Model validation — backtested on 49,256 real international matches
ballsignal-v1 has never been scored against reality. Here it is, against every recorded international result (1872–2026).
Calibration: model vs reality
ballsignal-v1 mean predicted probability vs the empirical World Cup base rate. Well-calibrated bars match.
Draw rate is stable — the model isn’t
Real draw rate by decade (≈22–23%, rock-steady) against ballsignal-v1’s mean predicted draw share.
Goal-margin distribution
How matches actually finish (49k results). Mean 2.94 goals/game (1.76 home / 1.18 away). The draw column is the only outcome the model over-weights.