STATFEED SCORE
A 0–10 rating for every match you play — computed on your PC, from the game's own Stats API events, weighted by when things happened, and neutral to the final score.
This page opens with what it can't see — the blind spots first, because that's what makes the rest worth trusting.
New here? Read the plain-language “what is the Statfeed Score” explainer. Curious how it stacks up against other tools? See the comparison.
What the score does not see
The Stats API reports events — goals, saves, shots, touches. It does not report positioning. Everything below is invisible to the score, permanently or until proven measurable:
- ×Rotation and positioning. Being exactly where your team needed you, so nothing happened — the best defense often produces zero events.
- ×Shadow defense. The slow retreat that forces a bad shot is worth more than many saves, and generates no event at all.
- ×Boost economy. Starving an opponent of boost wins games. The API doesn't report it in play.
- ×The pass before the pass. Only goals, assists and shots are credited today. The touch that created the chance your teammate missed is on our measurement list (with an evidence gate), not in the score.
Rule of the house: nothing joins the score without measured evidence that it predicts goals. Signals in testing show up as stats and badges first — never as silent points.
When the game crowned one player and the Statfeed Score picked another
Match #204, a 1–3 loss, real data. The game's MVP crown went to the hat-trick. The Statfeed Score went to the player who kept the game from ending earlier. Both are defensible — ours is just explicit about why:
- A hat-trick, and the highest in-game score on the field
- Two of the three goals came with the result already decided
- One of the saves an epic save, plus a Savior and a demo
- Every one of those events happened while the match was still alive
The in-game score rewards volume; the Statfeed Score weighs when and against what scoreboard. In 87.5% of matches with a recorded crown, the two agree (chance would be 50%). The 12.5% where they don't is not an error rate — it's the rating having an opinion, with a receipt.
What 6.5 means — and to whom
The scale is anchored to real percentiles of a declared population, not to a feeling. 6.5 = the median performance of the calibration base. Half of all real performances land below it. A 6.4 is not a bad grade — it's the middle of the pack, and the app now says so next to the number.
The declared population: 1,084 player-performances across 271 matches — all 2v2, one player's lobbies, one skill region. That's the footnote of a local-first tool: percentiles compare you to this base, and the base is published with every number. Above 9.0 the scale is the least-measured part (few real performances up there yet) — the app marks those scores accordingly. 3v3 gets its own anchors before the scale ever claims to compare across modes.
What moves the number
Events, from the game itself
Goals, assists, saves, epic saves, shots, demolitions — recorded by Rocket League's official Stats API as they happen. The score's only negative is an own goal, and since v2.0 it costs a fixed penalty: the model refuses to guess whether your clear under pressure deserved context punishment.
Context leverage
The same save is worth more at 2–2 with 20 seconds left than at 5–0. Every event is weighted by the scoreboard and the clock at that moment, capped at 3× so no single moment devours the match. Goals scored after the game is decided are worth less — and the match report says so, in words.
Per minute of play
The total is divided by minutes of game clock, not wall clock — replays and kickoff pauses don't dilute anyone's rating (they used to; that fix is what v2.0 is). Then the result maps onto the 0–10 scale through the published percentile anchors.
What never moves it
- ×The final score. No win bonus. Correlation with winning has to emerge from play — and it does (see below).
- ×Style. Flip resets and ceiling shots earn badges, never points.
- ×Your teammate's finishing. Credit is for the chance created, not for whether it was converted.
We publish the invariants (this page) and the changelog (below). We deliberately don't publish the per-event price list as a farming menu — but every match you play exports a full receipt with the exact parameters of its version, so anyone can redo the math on real data.
Measured against reality, in numbers
A rating asking for respect should publish its own audit. Measured on the full local base (271 matches, remeasured under SFS-2.1, Aug 2026):
The fine print that makes it citable
- Frozen history
- Your score is computed and frozen at the final whistle, stamped with the model version. Recalibrations apply from that day forward — a score you already have never silently changes: when a version recomputes history, it says so here and in the changelog.
- Version changelog
- SFS-2.2 (Sep 2026) — center, clear and 1st touch now score (weights 6 / 12 / 3, fixed multiplier 1.0); anchors recalibrated on the real base (1,594 samples). By the rule it's a patch (rank correlation 0.9999; p95 of the score change 0.00). History was recomputed under this version — announced, not silent: most older scores move one digit, 0.1 to 0.3; the best-score record kept its owner. · SFS-2.1 (Aug 2026) — two event-order fixes: assist now inherits the following goal's context multiplier instead of the preceding one, and a goal's scoreboard state is captured before the goal lands, not after. The scale itself is untouched — a patch, not a new anchor set (rank correlation 0.999 against v2.0). History was recomputed under this version: the earlier numbers were a measurement bug, not last month's method. · SFS-2.0 (Aug 2026) — game-clock denominator; fixed own-goal penalty; anchors recalibrated on 1,075 samples. By our own versioning rule (rank correlation 0.945 against v1), that's a major version — not a quiet patch. · SFS-1.0-beta (Jul 2026) — first calibrated release.
- The receipt
- Every match exports a JSON receipt: the events that entered the score, the scoreboard and clock context of each one, the divisor used, the full parameters of that version, and the resulting ratings. Disagree with a number? Redo the math — the dispute becomes about a listed fact, not about a black box.
- What doesn't count as a match
- Ties broken by leaving, sub-minute fragments and abandoned games never enter the scale. Forfeits are stamped as such. Alternate modes (Dropshot, Hoops, Rumble…) are outside the scored population today — declared as a decision, not left as a bug.
Built so the number can be checked, not believed. If you run Statfeed and want to argue with a score — good. Export the receipt and argue with the listed facts.