Every prediction we make, graded in public
Onside v0.5 calls a player’s gameweek score to within 1.57 points on average — closer than FPL’s own projection (1.86) on the identical 25,518 predictions, none of which the model had seen before. Written down before kickoff, never revised after.
For the stats people · “typical miss” is mean absolute error (MAE) — every prediction counts, so it cannot be gamed by only calling the easy ones.
| Model | Avg miss (pts) | |
|---|---|---|
| Onside v5 | 1.570 | This model |
| FPL's own ep_next | 1.860 | The official projection |
| Onside v2 (closed form) | 1.900 | Our previous engine |
| Form, last 4 GWs | 2.320 | Naive baseline |
| Season mean | 2.240 | Naive baseline |
| Onside v1 (legacy) | 1.208 | Where we started |
Beat FPL's own projection in 8 of 9 graded gameweeks, and the naive baselines — recent form, season average — in just as many. Same matches, same players, no cherry-picking.
The calls that win you gameweeks
When a player goes big — a goal, an assist, a clean-sheet stack — our number was 1.5 points closer than FPL's own projection, across 1,630 haul performances. The hardest call in FPL is the one we win by the most.
The quiet 3–4 point contributors who keep a rank ticking over: 1.7 points closer than FPL over 1,126 of them.
Knowing when a player will do nothing is how you pick the right captain and bench. We were 0.4 points closer on 4,419 blank performances.
The full four-way cut — including the one band FPL’s zero-happy projection wins — is on the calibration page. Receipts keep their losses.
Gameweek by gameweek — every week on the board
Each row is one gameweek: how far our numbers landed from what players actually scored, next to FPL’s own projection over the identical players. Onside has been the closer model in 2 of 3 graded gameweeks so far — and the weeks we lost sit in the same table, because a log you can cite has to keep them. Only players who reached 60 minutes are scored, so nobody gets credit for predicting zero for a bench that never played.
| GW | Frozen | Graded players | Our avg miss | FPL avg miss | Closer model | Our best call |
|---|---|---|---|---|---|---|
| GW3 | 28 Aug | 211 | 2.468 | 2.770 | Onside v5 | Guéhi — said 8.6, scored 8 |
| GW2 | 21 Aug | 204 | 2.437 | 2.315 | Official FPL ep_next | Roefs — said 3.0, scored 6 |
| GW1 | 21 Aug | 210 | 2.551 | 2.595 | Onside v5 | Havertz — said 3.6, scored 6 |
For the stats people · “avg miss” is mean absolute error (MAE), points per player per gameweek. Our worst GW3 call: Bogle (said 1.8, scored 14) · Our worst GW2 call: B.Fernandes (said 1.9, scored 23) · Our worst GW1 call: De Cuyper (said 3.9, scored 17)
GW4 is already frozen — , before its deadline — and will be graded here whichever way it goes. The projections behind it are downloadable now, results column empty, from the open dataset.
What this number is, precisely
Accuracy claims are only worth as much as their definition, so here is ours. MAE is measured out-of-sample — the model never saw these gameweeks during training. Predictions are timestamped before kickoff and are never revised afterwards. The figure above is the validated-population measure; the ledger behind it is public, and the methodology page sets out how the grading works.
One honest caveat about the comparison: FPL’s ep_next has information advantages ours does not, so treat 1.86 as a reference point rather than a defeated opponent.
Before the Premier League: the World Cup, graded live
The same engine family called every 2026 World Cup match in public, timestamped before kickoff — 66% of results called correctly with draws counted against us (68 of 103), and 25 of 31 knockout calls. Every hit and every miss is still on the board.
THE FULL WORLD CUP RECORD →Questions
How accurate are Onside FPL predictions?
Onside v0.5 scores a mean absolute error of 1.57 points per player per gameweek across 25,518 out-of-sample predictions. For comparison, FPL's own ep_next scores 1.86 and a naive "recent form" baseline scores 2.32. Lower is better.
What does MAE mean in FPL?
Mean absolute error is the average gap between a predicted score and the actual score, in FPL points. An MAE of 1.57 means the projection is typically within about a goal's worth of points of what actually happens — across every player, including the chaotic ones. It is the standard way to grade a points model, and it cannot be gamed by only predicting easy cases — every prediction counts. The number that matters most is the comparison: on the same matches, FPL's own projection misses by 1.86.
Which FPL prediction tool is the most accurate?
It is not currently possible to say, because almost nobody publishes a graded record. Competing tools advertise accuracy without a scoreboard behind it. Onside publishes this page and the full prediction ledger so the claim can be checked rather than taken on trust.
Is the accuracy figure independently verifiable?
The predictions are timestamped before kickoff and the results are public, so the record can be recomputed by anyone. Predictions are never revised after the fact.