AI Fantasy Football: How AI Predictions Work, and What They Cannot Do

Search for "AI fantasy football" and you will find a lot of pages promising that an artificial intelligence will pick your team. Almost none of them tell you what the AI actually is, what it is reading, or whether it has ever been right. This page does, because the honest answer is more useful than the marketing one — and because it is the only way you can judge any of these tools, ours included.

"AI" in fantasy football usually means one of three very different things

They get the same label and they are not remotely the same product.

• A statistical projection model. It ingests match data, betting markets and team news, and outputs a number per player for the upcoming round — expected goals, projected minutes, predicted points. It does not write, chat or explain. It is the oldest and by far the most tested of the three.
• A large language model (a chatbot). It writes fluent fantasy advice. It is genuinely good at explaining concepts and summarising news, and genuinely bad at arithmetic it has not been handed — ask a general chatbot for next week's predicted points and it will produce a confident number it has no way to compute.
• An optimiser. Given a projection and your budget, it solves for the best legal squad. This is not really AI at all — it is linear programming, it has been solved for decades, and it is completely dependent on the quality of the projection you feed it.

Most of what wins fantasy football is the first and the third. The second is a writing tool.

What our model actually is

TheFantasyTool is a statistical projection engine, not a language model. There is no chatbot picking your team. What runs is a chain of models that has been built and refined over eight years, and every stage of it is a number with a source:

• Match outcome. Team-level expected goals for and against each fixture, anchored to the bookmaker market — the market is the single best-calibrated forecast of a football match that exists, so we start from it rather than pretending to beat it from scratch.
• Minutes. A projection of how long each player is likely to be on the pitch, which is the input that moves fantasy points more than any other. It is built from rotation history, role, and our predicted lineups, which are themselves scored against what actually happened — 85.8% of predicted starters across 1,886 scored lineups in the window to 2026-09-07.
• Player output. Expected goals, expected assists, clean-sheet probability, cards, and the defensive actions the newer scoring systems reward — each projected per player, per fixture, from that player's own underlying numbers rather than from his recent points.
• Points. Those projections are then run through the specific scoring rules of the game you are playing, which is why the same player carries different predicted points in FPL, UCL Fantasy and a holdet.dk game.

The full chain is written up on our methodology page.

Why we publish our accuracy, and why you should demand it

A projection that is never scored is just an opinion with a decimal point. Ours is scored: model performance shows how our predicted points compare with what players actually scored, so you can see where the model is strong and where it is not.

This is the single most useful question to ask of any AI fantasy tool, and it filters the field fast: show me your predictions from before the round, scored against what happened. Not a testimonial, not a screenshot of one good week — the whole distribution. A tool that cannot produce that has not measured itself, and a model nobody has measured is a model nobody has corrected.

What an AI projection can and cannot do

It is worth being precise about the limits, because the failure mode of these tools is a user who trusts the number further than it deserves.

It can compare several hundred players on one consistent scale, which no human does reliably; price in a fixture swing or an odds move within minutes; and stop you making the two most common mistakes in fantasy — chasing last week's points, and valuing a player on his goals rather than on the chances he is actually getting.

It cannot know something that has not been reported. A manager's press conference an hour before the deadline, a knock in training that nobody has written about, a rotation decision taken on the day — the model sees these when the world does, and not before. It also cannot tell you how much risk you should take: a projection is an average, and whether you want the safe 5.2 or the volatile 5.0 depends on your rank and your league, which is a judgement about you rather than about the player.

And no projection is ever a certainty. A predicted-points figure of 6.4 does not mean he will score 6.4; it means that across many repetitions of that fixture, 6.4 is the average. Fantasy football is a game of small samples, and a good model loses plenty of individual weeks.

How to use it without switching your brain off

• Use the projection to build the shortlist, not to make the final call. Let it rank the pool; you decide between the top few on the things it cannot see.
• Look at projected minutes before predicted points. A big projection built on uncertain minutes is a bigger risk than the single number suggests.
• Check the ownership alongside it. The same projection means something different for a template player than for a differential, because rank is relative.
• Treat the optimiser as an argument, not an instruction. The optimal squad is optimal given the projection. When it suggests something that looks wrong, the interesting question is which input it is leaning on.
• Re-check close to the deadline. Odds move, lineups get confirmed, and the numbers move with them.

Where to start

The projections are live and most of them are free to read:

• Best FPL picks by predicted points — the full pool ranked for the upcoming gameweek.
• Optimal team calculator — the squad the solver builds from those projections.
• Predicted lineups — who we expect to start, and how confident we are.
• All our FPL tools in one place.

And if you want to know how any of it is calculated before you trust it, start with the methodology and then go and check the scoreboard. That order — how it works, then how well it works — is the whole point.