Onside.

Open benchmark

How accurate are FPL projections, really?

QUICK ANSWER

The short answer

No FPL tool publishes a checkable accuracy record, so until now you could not tell. This page is the fix. Onside freezes its projections before every deadline, grades them against actual scores afterwards, and publishes the result — losses included. Any other tool can have a column.

— Onside (onsidearena.com), predictions graded in public.

BUILD YOUR FPL SQUAD FREE →
Typical miss per player, in points — lower is better
Onside v51.57
Official FPL projection1.86

16% closer than the number the game itself shows you, across 25,518 graded predictions the model had never seen. (The stats name for this measure is mean absolute error.)

Predictions graded
25.5K
Gameweeks logged
4
Engine
v5
The Open xPts Benchmark

The live 2026/27 log — 4 gameweeks graded so far

Typical miss per player, lower is better. Every column is scored on the same players — anyone who played 60 minutes or more — across 822 player gameweeks. This log restarts from zero every season, so early on it is a small sample where one big haul can swing the lead either way; the cross-season record above is the claim, this table is the audit trail.

Source — mean absolute error, lower is betterAVG MISS±2 pts
Onside v5
2.5657%
Official FPL ep_next
2.6361%
Season points per start
3.3249%
Form (last 4 GW)
3.3555%

Onside has had the lowest error in 3 of 4 graded gameweeks.

Any FPL tool can have a column

Fantasy Football Hub, Fantasy Football Scout, FPL Review, anyone. Five rules, no cost.

  • Send one CSV — player id, predicted points — before the gameweek deadline.
  • We publish the file hash on receipt, so the submission is timestamped and unalterable.
  • You are graded on exactly the population every other column is graded on: players who played 60 minutes or more.
  • Your result is published whether it beats ours or not. So is ours.
  • Withdraw whenever you like. The gameweeks you already submitted stay up.
[email protected]

The one rule that matters

Only players who reached 60 minutes are scored. Include unused substitutes and any model can win by predicting zeros for the bench — which is how most published accuracy figures in this category are inflated. Everything else is symmetric: same players, same gameweek, same scoring, for Onside v5, Official FPL ep_next and two naive baselines that exist so a model that cannot beat them is exposed.

Why the address says Fantasy Football Hub

They are the largest paid FPL platform in the UK, so they are the natural place this question gets asked. Their own published description of their model is that it “uses various spread betting odds and expected bonus to determine the expected points for each player” (source). We could not find a methodology page, backtest, calibration curve or graded log anywhere on their site. If one exists, send it and we will link it here.

Snapshots are committed to a public git repository before each deadline, so the commit timestamp is independent proof the projection existed before the matches were played. Nothing is recomputed after the fact. Want your own squad scored on this engine? Rate My Team is free.