Is FPL AI accurate? Every prediction we made, checked against real points

The short answer

Through the first six gameweeks of 2026/27, our predictions missed a player's real FPL score by about 1.7 points on average (gameweek errors between 1.55 and 1.77). That is across roughly 450 to 500 players every week, not a hand-picked shortlist.

FPL scoring is very random — one goal is worth 4 to 6 points — so no model can call individual hauls. What a good model can do is rank players correctly on average, so that over a season your decisions are right more often than wrong.

Gameweek by gameweek (2026/27)

Every past gameweek page (for example /gameweek/6) shows the predictions we published before the deadline next to the real points, unchanged. The home page keeps the latest numbers up to date.

• —GW1 — 500 players, average miss 1.70 points
• —GW2 — 469 players, average miss 1.55 points
• —GW3 — 466 players, average miss 1.64 points
• —GW4 — 453 players, average miss 1.77 points
• —GW5 — 441 players, average miss 1.74 points
• —GW6 — 431 players, average miss 1.68 points

Where we are wrong

So far we have slightly over-predicted on average — by between 0.2 and 0.4 points per player per week. The biggest misses come from minutes: surprise benchings and early substitutions. We report this openly rather than hiding it, and the model is re-calibrated as each gameweek's results come in.

Last season's replay

We also replayed the whole of 2025/26 week by week, using only data that existed before each deadline: £100m budget, one free transfer a week, no hits. The base replay and the chip-adjusted total are shown on our home page, next to the hindsight ceiling that no real manager could reach.

How to judge any FPL AI tool

• —Does it publish its predictions before the deadline, so they can't be edited later?
• —Does it score every player, or only show its best calls?
• —Does it say how big its average miss is, and in which direction?
• —Does it respect FPL rules — budget, three players per club, free transfers? General chatbots often don't.