How SquadPilot's FPL projections work.

Expected minutes are never a guess at one number. A player has two routes onto the pitch, and they are worth very different amounts:

So four separate things have to be estimated, and you see all four:

• Start probability: how likely this player is to be in the XI.
• Minutes if they start. A 90-minute centre-back and a 60-minute rotation forward are not the same asset.
• Cameo probability and cameo minutes: what a bench role is actually worth.
• Rotation risk and role confidence: how far the model trusts its own read.

What feeds it

• A persisted, structured role prior for each player: nailed starter, rotation risk, impact substitute, backup.
• Historical starts and minutes.
• Current-season evidence, weighted in as it accumulates through the campaign.
• Official FPL availability flags. Injury status and chance-of-playing percentages, straight from the official API.

Where it is honest about not knowing

Every expected-minutes figure carries an evidence freshness stamp. Early in a season, with no current-season minutes to lean on, a player's estimate is marked prior only. That is the model telling you it is running on history rather than observation. Manual role priors and minute overrides stay explicit and reversible, never silently baked in.

It is also the answer to the classic August failure mode. Before a ball is kicked, naive rate-based models will happily rank a backup goalkeeper above a first-choice striker. Modelling minutes first, and labelling low-evidence reads as such, is how that gets caught.

What else the minutes model produces

These views reuse the same start probabilities rather than re-deriving them, so they can never disagree with the projections underneath:

• A predicted starting XI for every club. The eleven most likely starters, constrained to a legal formation, with a confidence read off the tightest within-position margin and every contested place named: who is in, who is pushing, and the gap between them.
• Set-piece order per club. Penalties, corners and free kicks, in order, with how long each duty has been held. A taker promoted last week and one who has had the ball all season are different bets.
• A claim in the source ledger. Before each deadline the predicted XIs are filed under the source name squadpilot-engine and graded against the line-ups that actually took the field, by the same code that grades every external team-news source. We do not get to mark ourselves by a kinder rule than we mark them.

Expected points are projected per fixture and then summed, rather than per Gameweek. Double Gameweeks and blanks then fall out of the arithmetic on their own, with no special case to maintain.

Each fixture projection combines:

• the expected minutes distribution from step one;
• persisted expected-goals and expected-assists rates (raw goals and assists only as an explicit, labelled fallback);
• official home/away and fixture difficulty;
• stored team strength fields where they are usable;
• the clean-sheet and goals-conceded picture for that specific fixture, modelled from how many goals each side is expected to score;
• goalkeeper saves;
• historical bonus rate;
• the odds of clearing the defensive-contribution threshold, scaled to how many minutes he is expected to play.

Every number can be taken apart

A projection you cannot interrogate is worth very little. Each result carries its component values, the fixture it belongs to, the model version and a calculation timestamp. Missing data gets named rather than papered over. An absent fixture feed is reported as fixture data unavailable, and a team proven to have no fixture in a populated Gameweek is reported as no fixture, never as a quiet zero.

A transfer is evaluated against the squad you actually have, under the rules that actually bind:

• Expected-points gain over a horizon, not just the next Gameweek, because a transfer you make now constrains the squad you have in five weeks.
• The price of the move and what it leaves in the bank.
• The points hit, if the move needs one, subtracted from the gain rather than mentioned separately.
• Squad legality: budget, the three-per-club limit, formation and position quotas.
• Risk on both sides: start probability, fitness flags, rotation risk and fixture difficulty for the player coming in and the one going out.

The four parts you always get

1. Action. The named move, with the deadline attached.
2. Expected. The projected gain, and the expected minutes on each side that produce it.
3. Risk. The case against, stated plainly, including what it costs you if the call is wrong.
4. Alternative. The other defensible path, usually holding and banking the free transfer, scored on the same terms.

The alternative earns its place. Most weeks you will want to weigh the recommended move against doing nothing, and that case is scored on exactly the same terms.

Thirty minutes before the deadline the inputs are refreshed, the model runs, and the projection is written to an append-only ledger with a timestamp, the ingest snapshot it came from, and the model version that produced it. After the final fixture, realised outcomes are matched against it and the error is published, misses included, alongside how the model compared against the free baseline FPL already gives you.

Read how the public scorecard is measured →