FPL AI · THE RECEIPTS

How accurate are AI FPL predictions?

Most AI FPL tools ask you to take accuracy on faith. This page does the opposite: the actual numbers, the benchmarks they beat, the public record they were graded against — and the honest list of what no model can predict.

QUICK ANSWER

The short answer

Onside's AI predicts FPL points with a mean absolute error of 0.86 per player per gameweek — measured on 51,518 predictions the model had never seen, beating FPL's own expected-points figure (0.896). The same engine called 81.9% of decided World Cup 2026 matches, graded in public.

What MAE means, in plain English

MAE — mean absolute error — is the average gap between a predicted score and the real one. If the model says Saka will score 6.2 points and he scores 8, that's an error of 1.8; average those gaps over every player and every gameweek and you get the MAE. Lower is better. An MAE of 0.86 means the typical prediction lands within about 0.86 points of reality — and crucially, it's measured out-of-sample: on gameweeks the model was never trained on, which is the only honest way to score a predictor.

The benchmarks

0.86
Onside (current engine)
Cross-season, out-of-sample
THIS SITE
0.896
FPL's own ep_next
The game’s built-in expected points
1.025
Onside previous engine
Our earlier version, for context
1.048
Form (last 4 GWs)
The naive "pick on form" rule
1.054
Season average
Assume every player repeats his mean

MAE in FPL points per player per gameweek, lower is better. All figures measured on the same out-of-sample gameweeks (51,518 predictions across two seasons). ep_next is FPL's own expected-points column — the benchmark any serious model has to beat.

GRADED IN PUBLIC

The World Cup 2026 stress test

Backtests can flatter. So through World Cup 2026 the same engine named a favourite before every match and was graded after full time, in public, with the misses left on the board: 68 of 83 decided matches called correctly — a 81.9% hit-rate across 103 graded matches. No cherry-picking; every played match is on the record.

See every graded call →

Calibration: when we say 60%, it lands about 60%

Accuracy alone isn't enough — a useful predictor also has to mean what it says. A model is calibrated when its confidence matches reality: of all the calls it makes at 60% confidence, about 60% should come true. Onside publishes its calibration openly, banded by confidence level, so you can check that an "80% captain call" really behaves like one. That's the difference between a probability you can plan transfers around and a vibe with a number attached.

The calibration page →
THE HONEST BIT

What AI can't predict

No model — ours included — can predict a red card, a hamstring in the warm-up, a manager's surprise rotation, or a deflection off a defender's knee. Single gameweeks are dominated by exactly this kind of noise, which is why even the best predictor will look wrong on any given Saturday. What AI genuinely offers is different: consistently better-than-human odds, compounded over 38 gameweeks. Judge any FPL AI on its season-long, out-of-sample record — never on one week's haul or horror show. If a tool won't show you that record, that tells you something too.

Accuracy FAQ

How accurate are AI FPL predictions?

The honest benchmark is mean absolute error (MAE) — how far the predicted points land from the real points, on average. Onside's engine scores 0.86 per player per gameweek across 51,518 out-of-sample predictions spanning two seasons, against 0.896 for FPL's own ep_next figure and 1.048 for a naive last-4-gameweeks form rule. Most AI FPL tools publish no accuracy number at all, so treat any unverifiable claim with caution.

What does an MAE of 0.86 actually mean?

On average, Onside's predicted score for a player lands within about 0.86 points of what he actually scores that gameweek. Football is high-variance — a last-minute penalty or a red card swings any single prediction — so no model gets individual weeks "right" every time. What a lower MAE buys you is a consistent edge across a 38-gameweek season: better captain calls and better transfer priorities on aggregate, which is where FPL ranks are actually won.

Has the accuracy been tested on real matches?

Yes, twice over. The FPL MAE is measured out-of-sample — on gameweeks the model had not seen — across 51,518 predictions. And at World Cup 2026 the same engine's match calls were graded in public after every full-time whistle: 68 of 83 decided matches called correctly (81.9%), with every miss left on the scoreboard at onsidearena.com/world-cup-2026/model-record.

What can't AI predict in FPL?

Red cards, injuries mid-match, surprise rotation, penalty shouts and deflected goals — the one-off events that decide single gameweeks. AI predictions are probabilities, not certainties: they tell you which choices are most likely to pay off over time, not what will happen on Saturday. Any tool claiming to "know" who will haul this week is overselling. The realistic promise is a measurable edge across a season, which is what the graded numbers on this page show.

← FPL AI hub
Everything the AI does for FPL
ChatGPT vs a purpose-built model →
Where each one earns its place
Best AI FPL tools 26/27 →
The honest shortlist
FPL 2026/27 · SEASON STARTS 21 AUG

Everything for your FPL 26/27 season

The same model that graded every World Cup call in public, pointed at Fantasy Premier League.