How the PitchVAR rankings work

Every number on our rankings table comes from one deterministic replay of every recorded international since 1872 — and every change to the formula must beat the current one in a measured test before it ships. No committee, no vibes, no exceptions.

The replay

Each side starts at 1500. Every match moves both ratings by an amount that depends on the result, the expected result (a logistic curve on the rating gap, with a home-advantage bump on non-neutral venues), the margin (diminishing returns past two goals), and the competition (a World Cup match moves ratings twice as hard as a friendly). The replay runs chronologically over the full spine — 48,000+ matches — so a rating is never an opinion about a team; it is the compressed record of everything that team has ever done.

The uncertainty axis

A team that has not played for years is a question mark, and the system treats it like one: after a year of silence a side's rating moves toward its results up to twice as fast, scaling with the length of the gap. This is the one recent change to the formula — adopted 2026-07-28 after it beat the previous engine across every era of historyin our acceptance test (paired log-loss on 48,137 matches, t = 15.1).

The gate: how formula changes ship

Any proposed change is replayed over the whole spine next to the current engine, and each match's pre-match probabilities are graded against what actually happened (log-loss). The challenger ships only if it predicts better overall and in the most recent era. The gate rejects our own ideas more often than not — recently it threw out shootout-winner bonuses (a shootout is close to a coin flip; paying rating points for winning one adds noise) and knockout stage weightings (a final carries more stakes, not more information — weighting it measurably hurt predictions). What you see is the best formula we have been able to prove, not the cleverest one we could imagine.

What a rating refuses to say

Only sides that have played within the last four years — one World Cup cycle — are ranked; historical entities keep their pages but not a seat at the current table. Matches decided in extra time grade on the final score; shootouts grade as draws, because that is what the pitch said. And where the record itself is ambiguous, we show an em-dash rather than a guess — wrong data is a scandal, a gap is a shrug.

Rankings are not predictions

The match pages carry our model's pre-kickoff read for every fixture. Those predictions are graded publicly against results — and judged against the betting market's closing prices, which we treat as the examiner, never as an input. The rankings feed the model; the scoreboard keeps both honest.

Formula changes are dated and recorded. Current engine basis: full-history replay, competition-weighted K in [0.7, 1.4], home advantage 60 (neutral-exempt), margin multiplier capped-growth, inactivity-scaled per-team K (1× → 2× across years 1–3 of silence), four-year ranking eligibility. Adopted 2026-07-28.