fpl model · ai team · pre-season

The Bawler AI team

Every other page here asks you to trust a projection. This one takes the same risk you do. One squad, picked and run entirely by the model for the whole season — its own transfers, its own captain, its own chips, decided by a fixed policy written down below. Every decision is published before the deadline and scored after it. Nothing is edited with hindsight, and the bad weeks stay on the page.

The squad as it stands

The violet number on each card is projected points per gameweek. An amber or red percentage means the lineup research rates that player a rotation risk — nailed starters carry no badge.

Haaland wears the armband. The rotation percentages come from our predicted-lineup research across all 20 clubs — it is what stops the model picking a projection that belongs to someone who does not play. The projection method and the full backtest are on predicted points.

Chips

All four are costed in projected points every single gameweek and played only when this week is the best week the model can see — if a bigger one is coming, the chip is held. They reset at Gameweek 20, so each is available twice: once in each half of the season.

Worth saying up front, because it is the least flattering thing we know about this part of the model: chips are the weakest thing it does. Played out over three past seasons they were worth roughly +21, +31 and +7 points in total, and the wildcard on its own was worth nothing at all. So the chips below are played by the rules, and the rules are published, but do not read a wildcard here as the model knowing something. The part that actually works is the weekly transfers.

The plan for all eight

There are no double or blank gameweeks to plan around yet — and that is not a gap in the model, it is the fixture list. Every club currently has exactly one match in every gameweek. Doubles and blanks only appear later, when matches are postponed for cup runs or weather and rescheduled. Naming a gameweek for a Bench Boost today would be inventing one, so the chips that need a double or a blank are held, with the trigger stated. This is rebuilt every night and will name gameweeks the moment the calendar creates them.

The decision log

Written by the model at the moment it decided, newest first. This is the part that makes the record worth anything: a transfer that looks obvious afterwards has a timestamp showing when it was actually made, and a bad one is still here.

The rules the model plays by

These are fixed in code and published so the season is a real test rather than a highlight reel. There is no human in the loop: nobody overrules the model on a hunch, and nobody quietly edits a pick that aged badly.

• squad: £100.0m, 2-5-5-3, max 3 per club; the 15 that maximises the projected XI, plus what the bench is genuinely worth — each substitute counted by the chance a starter fails to appear AND he does, in autosub order
• transfers: made by the same transfer engine the three-season backtest measured, not by a separate set of rules written for this page. Free transfers accumulate — one a gameweek, unused ones roll over to a maximum of five — and the engine spends as many as it holds, plus at most one -4 hit above 6. It weighs a whole week's moves together rather than one swap at a time, so it can sell one player to afford another. Each free transfer it spends has to be worth 2 points, counted as what the best starting XI plus the captain gains over the next 4 gameweeks, with the nearest week counting for most. A week where nothing clears the bar banks one instead. At five the bar halves, because a transfer that cannot roll further is lost if unused. Never for a player starting < 55% of matches or flagged ≤ 50%
• captain: in season, the highest CEILING in the XI rather than the highest average — the armband doubles one player for one week, so his chance of a big score is what matters, and captaining by ceiling beat captaining by mean 202-187 across the backtested season. Pre-season, before any gameweek exists, the armband follows the same opening-run projection the squad is picked on
• chips: every gameweek all four are costed in projected points and played only when this week is the best week in view: triple captain above 8, bench boost above 14, free hit above 12, wildcard above 15 across the horizon. A chip is held if a later gameweek in the horizon looks better, and the bar drops as the window closes. Each is available once per half (they reset at GW20), one chip per gameweek; a wildcard replaces that week's transfers and a free hit changes only that week's team sheet
• frozen: every decision is written here before the deadline and never edited after it. A gameweek whose deadline passed before a decision could be recorded is logged as missed and left alone — no transfer, no chip, no team change — rather than filled in afterwards

What this model is bad at

Here is what we already know it gets wrong, said before the season can say it for us. The same code played three full past seasons from start to finish, making every call with only the information available at the time, and it was not good at everything.

• Picking the opening squad is its weakest skill. Its own Gameweek 1 fifteen, left untouched for a whole season, finished below the average manager in all three seasons tested. The squad on this page today is the part to trust least; what it does with it from September is the part that counts.
• Chips are close to worthless. All eight plays together came to about +21, +31 and +7 points across those seasons, and the wildcard specifically to nothing at all. It still plays them by the published rules, because a rule you follow when it is unflattering is the only kind worth publishing.
• It needs the season under way before it is any good. Its edge in the first half of a season was much smaller and much less reliable than in the second, because everything it knows comes from matches that have actually been played.
• What it is genuinely good at is the weekly grind. The transfers and the team it picks each week were worth +330, +511 and +537 points over those three seasons against leaving the opening squad alone. That is the whole case for this team, and everything above is the cost of making it honestly.

Why we are doing this in public

A projection model is easy to market and hard to hold to account. Over a season this squad produces a number nobody can argue with: what it actually scored, against what the average manager scored, from decisions that were all published in advance. If the model is good you will see it here; if a rule in the policy is wrong, a bad run will show that too, and the fix will be a change to the rules with a date on it.

Want to play alongside it rather than copy it? Load your own 15 into the team builder, or import your real side on my team and compare the two.

Computed from the Official FPL API — pure fantasy-points analytics, no betting angle. Projections are model estimates, not guarantees.