How my FPL predictor works (and how I use it)

I built a little Fantasy Premier League predictor that lives at fpl.mincho.dev. It picks an “optimal” squad for the upcoming gameweek, and — once the season starts — grades itself against reality. This is a quick tour of what’s under the hood, because the whole point of it is that it’s not a black box.

Where the data comes from

Everything comes from the official Fantasy Premier League API — the same public data the game itself uses. No scraping, no paid feeds. For every player it reads their expected goals and assists, expected goals conceded, defensive actions, minutes, availability, price and position; plus each team’s fixture difficulty.

How it scores a player

For each player it estimates expected points for the gameweek, built from parts you can actually reason about:

• Attack — expected goals × the points a goal is worth for that position (a midfielder’s goal is 5 points, a forward’s 4), plus expected assists.
• Clean sheets — the chance a team keeps a clean sheet, from expected goals conceded (defenders and keepers get 4 points for one).
• Defensive contribution — the newer points for tackles, blocks and interceptions.
• Appearance & bonus — the reliable floor of just playing, plus a bonus estimate.
• Fixtures — all of the above is nudged up for easy upcoming opponents and down for hard ones, blended over the next four gameweeks so a good run counts, not just the next match.

A couple of statistical touches keep it honest: rare events (clean sheets, hitting the defensive-contribution threshold) use a Poisson probability rather than a naive average, and lightly-played players get regressed toward the average so a small sample doesn’t make someone look like a superstar.

Picking the team

Given every player’s projected points, an optimiser then chooses the best legal 15-man squad — within the £100m budget, the 2/5/5/3 position split, and max three per club — and the best starting XI and captain. It’s not “pick the 11 highest scorers”; it’s the best value team the rules allow, which is why a £15.5m striker sometimes gets left out for two players who together score more.

How often it updates

It runs itself on a schedule (a free scheduled job), and it’s fixture-aware so it never jumps the gun:

• Before each deadline it rebuilds the recommended squad using the latest prices and news — only in the couple of days before the deadline, not weeks early.
• After each gameweek (once the points are final) it scores its own predictions and works out the dream team — the best possible XI in hindsight — so you can see how close the model got.

So the site is always current for the next deadline, with no button to press.

How I read it

• The pitch is the recommended squad and captain for the upcoming gameweek.
• The accuracy section (once the season is underway) shows predicted vs actual and a season trend — is the model getting better or worse over time?
• The dream team is a yardstick, not advice — by the time it exists the gameweek is over. It just shows the ceiling I was aiming at.

The honest caveats

It’s a v1 and it says so. Pre-season it runs on last season’s numbers, so early projections are rough — it sharpens itself as real games pile up. It doesn’t yet model penalties explicitly, and things like promoted-team players are probably a little over-rated until there’s real Premier League data to calibrate on. That’s fine: it’s a transparent model I can see being wrong and fix with evidence, which is the whole reason I built it this way.

More to come as the season starts and the predictions meet reality.