Methodology

How FPL Tactics Works

Every projection on this site comes from one model with three moving parts: a Fantasy Rating, an xGI ensemble, and a multi-gameweek MILP transfer planner. The Fantasy Rating is built from five Elo components and Game Time, a read of the minutes a player is getting. This page walks through what each part does, how we validate it, and why we made the calls we made.

The projection model, end to end

For each (player, gameweek) pair the system yields one number: expected fantasy points. At its core is a blend of two predictors that work independently, each with its own model and its own data. Three more steps then shape it: how many minutes he is likely to play, a learned model that takes over the next gameweek, and FPL’s own injury and availability flags.

The Elo predictor keeps a set of ratings per player, one for each way of scoring, and moves each one match by match by the gap between what he earned that way and what his rating expected. Each position has its own ceiling, so a keeper’s clean-sheet run and a forward’s hat-trick are judged on scales that fit them. Goalkeepers are the exception: a keeper is projected from one overall rating, measured against the keeper ceiling, because splitting keepers into parts added noise without making them more accurate.

The xGI predictor tracks each player’s expected goals and expected assists per 90 minutes - the chances he is getting, whether or not they went in. A new signing with three good matches would otherwise look like a world-beater, so we pull every player’s number partway toward the positional average until he has enough minutes to stand on his own.

We don’t blend the two predictors with one global weight, because the right mix of xGI and Elo depends on context - xGI tells you a lot about a premium forward at home, much less about a defender. So, for goals and for assists separately, every outfield player-match falls into a slice by position group (midfielders and forwards together, or defenders), home or away, and premium or not, and each slice gets its own weight α. Fixture difficulty is left out of the slices on purpose: it is applied once, to the finished projection (see below). The fit is ridge regression against realised points, anchored to a 50/50 blend so slices with thin data don’t overfit to a handful of noisy examples.

On top of the goal, assist, clean-sheet and defensive-action parts, each projection adds appearance points (see Game Time below) and an allowance for what those parts don’t cover, mostly bonus points, from his last five matches and pulled toward his position’s average while that record is short. The total is then scaled by fixture difficulty. The factor comes from how the same player has scored against each of FPL’s five difficulty levels over every past season we hold, worked out separately for each position: a defender’s points fall further against a top side than a midfielder’s, because clean sheets go first.

The next gameweek gets a second opinion. A next-week model, trained on past seasons, reads his recent form and minutes, last season’s record, his club’s recent results and the fixture, for the 300 players with the most points so far. It supplies the whole projection for the next gameweek, half of it for the one after, and none beyond that: one week out it ranks players better than the main model, six weeks out it ranks them worse.

Last come FPL’s flags. An injured or suspended player is scaled by his chance of playing until the gameweek his FPL news says he is back, and in every week when it gives no date. A doubtful regular starter is cut for the round the flag is about and eased back over the next few, at the rate doubtful regulars have actually come back. Any other doubtful player is scaled by his chance of playing in every week.

What “premium” means

A two-way flag. A player is premium when his Finisher, Playmaker and Clean Sheet ratings added together sit in the top quarter of the training data, and non-premium otherwise. The cut-off is worked out on the training seasons and then frozen. The most expensive forwards and midfielders sit in the premium slice; the average rotation player does not. We split on it because the best mix of xGI and Elo really is different for the two groups.

Why two predictors and not one

Elo captures what xG alone misses: clean sheets, bonus points, defensive contributions. xGI catches attacking returns that Elo is slow to react to after a hot streak. On seasons it was never trained on, the ensemble’s typical miss is 5–6% smaller than Elo alone, and it guesses low by 29–46% less. The numbers are on the accuracy page.

The Fantasy Rating components

The player page shows up to six bars. Five are Elo ratings, each learned from one kind of FPL points. The sixth, Game Time, isn’t an Elo at all: it is the minutes the player has been getting.

  1. Finisher - learned from FPL goal points. Defenders score too rarely for a goal rating of their own, so in a defender’s projection goal threat comes from his overall rating instead.
  2. Playmaker - learned from FPL assist points.
  3. Clean Sheet - learned from clean sheets and goals conceded; for goalkeepers, defenders and midfielders.
  4. Shot Stopper - learned from save points; goalkeepers only.
  5. Defender - learned from FPL’s defensive contribution points, paid since 2025/26: 2 points in a match where a defender makes 10 or more clearances, blocks, interceptions and tackles, or a midfielder or forward makes 12 or more of those plus ball recoveries. Goalkeepers can’t earn them. Each match is judged on its own count, so a double gameweek is two chances, not one combined total. The rating starts at what each position actually earned per 90 minutes in 2025/26 (about 0.52 points for defenders, 0.33 for midfielders and 0.02 for forwards) and learns only from seasons with the rule. A match he plays under 21 minutes, or a week he doesn’t play, leaves it where it was: whether he plays is priced once, by his minutes, not taken off this rating as well.
  6. Game Time - not an Elo. His average minutes over his last five matches this season, out of 90, on a 0–100 scale: a 90-minute regular reads 100 and a striker averaging 45 minutes reads 50, however good his Finisher or Playmaker numbers are. Early in a season, the matches he hasn’t played yet are filled with his average minutes from last season’s appearances. A match he missed counts as zero here, so a player who is still out still reads as out.

The bars don’t each add a fixed share. The projection uses each rating where it applies to the position - Finisher and Playmaker weigh heavily on a forward, Clean Sheet and Defender on a defender. A goalkeeper is projected from one overall rating instead, so his Clean Sheet and Shot Stopper bars describe him but don’t drive his projection. Since November 2026 that rating is measured against a keeper’s own ceiling rather than a midfielder’s, which halved how far keepers were over-projected in our tests.

Game Time and the projection. Game Time scales the overall Fantasy Rating: two players with identical Finisher and Playmaker numbers rate very differently if one is locked in for 90 minutes a week and the other is splitting time with a competitor. The projections the planner uses go further than the five-match average, in three layers:

  • Expected minutes. Once a player has two gameweeks at his club, his expected minutes come from the competition for his place: his share of his position’s minutes at the club, adjusted when a regular team-mate is out and for doubles and blanks. Until then they come from the five-match average, with one difference from the bar: a match he missed while FPL had him flagged out or doubtful is left out rather than counted as a zero, so a player back from a knock is not held down by the game he sat out.
  • The learned minutes model. For outfield players, over the six gameweeks a plan looks at, a model trained on past seasons reads those expected minutes alongside his recent minutes, starts and substitute appearances, how many rivals he has for his place, his price and the fixtures. It gives the chance he doesn’t play, the chance he plays under an hour, the chance he plays an hour or more, and how long he is likely to be on. Appearance points then count at their average value instead of a flat 2 for anyone expected to play an hour: even a player expected to play the full 90 misses the whole match about one week in nine. Goalkeepers keep the expected minutes above, because the model pulled a near-certain 90 down and cost keepers accuracy.
  • FPL’s flags, applied last, as described above.

For the next gameweek, and half of the one after, most of the projection comes from the next-week model, which reads minutes its own way.

The practical effect: a defender heading into a soft run of fixtures can have his Clean Sheet and Defender numbers both climb on the back of clean sheets, his Finisher and Playmaker move (or not) on completely separate evidence, and his Game Time hold steady so long as he’s still nailed in the starting XI. Three independent stories, one number on the page.

A streak boost speeds up a rating’s updates for a player who started below the typical rating for his position and keeps outscoring it. This is what stops breakout players - Igor Thiago in 2025/26 is the textbook example - from being underrated for half a season.

The overall Fantasy Rating at the top of the player page is the whole model’s projection for one match against an average opponent, at his Game Time minutes, placed on a 0–100 scale against every player in the game. Each component bar is placed on its own 0–100 scale against today’s Premier League players at his position. The overall scale is fitted again at every refresh, the component scales whenever we refit the model.

The xGI ensemble

The xGI predictor turns one input - expected goal involvement per 90, shrunk toward the positional average - into expected fantasy points. The shrinkage keeps a new signing with three matches under his belt from being projected off a freak underlying run.

The weights are fitted by ridge regression on every past Premier League season we hold, from 2019/20, and never on the season being projected. FPL only started publishing per-match xG in 2022/23, so 2019/20 to 2021/22 use Understat’s figures; weeks where no xG was published (2022/23 gameweeks 1–15) count as unknown, not as zero. An α of 0.5 means a slice’s output is split evenly between xGI and Elo; the fitted weights vary from slice to slice.

The horizon planner

The transfer planner is a mixed-integer linear program (MILP), solved by CBC. It maximises projected points across a planning window - six gameweeks by default, each counting 85% of the one before - subject to every FPL rule that actually matters:

  • £100m budget, three-per-club limit, position quotas (2 GK / 5 DEF / 5 MID / 3 FWD).
  • Free transfers (one a week, up to five banked) and the −4 hit for each extra.
  • The half-rise sell mechanic: selling a player whose price has risen since you bought him only realises half the gain.
  • Per-gameweek decision variables: who to buy, who to sell, whether to take a hit, the starting XI and the captain.

We re-plan every gameweek. A plan that looked best at GW10 with a six-week view gets re-evaluated at GW11 against the new state - ratings have updated, prices have shifted, flags have changed. The planner doesn’t lock in later weeks’ moves; it acts on this week’s and asks the same question again next week.

Captain. Every page that picks a captain - the planner, Rate My Team, GW Tips and the planner pitch - uses one rule, and so do our backtests. For the next gameweek a captaincy model trained on past seasons scores the eleven the plan already chose, reading what the projection misses: form, ownership and transfer trends, penalty and set-piece duty, and the fixture. It picks the captain and vice-captain from that eleven and changes nothing else. Later weeks are captained by projection, because the model has only been tested one week ahead.

Chips (Wildcard, Free Hit, Bench Boost, Triple Captain) are planned around, not timed. Set a chip in a week and the planner plans that week with it: unlimited free moves on a Wildcard, a one-week squad on a Free Hit that reverts the week after, all 15 players scored on a Bench Boost, the captain scored three times on a Triple Captain. It only offers chips your team can still play. Choosing which week to play a chip is still yours. The accuracy page covers what this means for the validated results.

Plans worked out ahead. After each gameweek’s last match, once the new projections are made, the planner works out the default Steady plan in the background for each signed-in manager with an imported team who has planned in the last week or two, so pressing Generate can answer at once. Prices and injury flags are checked again before a stored plan is served.

The conviction threshold (T)

Every transfer the planner makes has to clear a bar set in projected points across the planning window: the plan with the move has to beat the plan without it by at least the bar. Wildcard and Free Hit weeks have no bar, and neither do moves you lock in yourself. The two strategies set the bar differently.

Climber: a flat T = 2.5. Every transfer must gain at least 2.5 points, and one that costs a −4 hit must gain the 4 as well. Before the November 2026 data rebuild a walk-forward sweep tried 2.0, 2.5, 3.0 and 3.5 across every strategy setting, and 2.5 landed in the top tier from both a strong and a deliberately weak squad. Re-tested on the rebuilt data over 2020/21 and 2021/22, Climber did worse at both 2.0 and 3.5.

Steady: a bar that falls as free transfers bank. Steady never takes a hit, so every move it makes spends a free transfer, and it saves them for the weeks they are worth most. The bar depends on how many free transfers the week starts with: 4 points with one (what a hit costs, since that transfer is the only cover against needing one later), then 3, 2 and 1, and 0 with five, where an unused transfer would be lost. So a quiet week banks its transfer, Steady makes several moves at once when they are worth more together, and a starter who stops playing is still replaced. Over the season’s last five gameweeks every bar moves step by step to Climber’s flat 2.5, reached at GW38, since a banked transfer has fewer weeks left to be used.

Steady’s plans also always consider every premium (£10.0m or more) and the best few players at each position priced for a big move, whatever their next fixtures, so the transfers it saves can reach the players worth saving for. A Steady plan gets 20 seconds to solve; with no plan by then, the planner falls back to a no-hits plan with the flat 2.5 bar and without those extra players.

From a good pre-season squad in our backtests, Climber made about 40 transfers a season, about 3 of them hits, and Steady about 36, none of them hits.

The two strategy archetypes

The planner ships with two presets. Both look six gameweeks ahead and use the same projections and the same captain rule. They differ in whether the planner may take a −4 hit, and in how much a transfer has to gain (see the conviction threshold).

Steady (default) - h = 6, no hits, banks free transfers

For the manager maintaining an established squad. Steady is never allowed to take a −4: every transfer it makes is funded by a free transfer, banked on the 4-3-2-1-0 schedule above.

On the rebuilt data, from 20 realistic pre-season squads over four test seasons (2022/23 to 2025/26), Steady averaged 2218 points a season as the live planner runs it, with no chips played. Like for like, Climber averaged 2190 against Steady’s 2212, a gap within the noise, so from a good squad the two finish level. Against the older Steady, which asked a flat 2.5 of every transfer and had no extra top players, the banked Steady scored 11 points a season more, also within the noise, so we count it as level rather than better. Pick Steady if you want the explicit no-hits guarantee; the accuracy page has the full results.

Climber - h = 6, T = 2.5, hits allowed (up to 2 per gameweek)

For the manager whose squad has work to do. Starting from a deliberately weak pre-season squad, Climber averaged 50 points a season more than Steady across four test seasons on the rebuilt data, and came out ahead on 14 of 20 squads.

The hit budget is self-regulating. Climber is allowed up to two −4 hits per gameweek, but the MILP only takes one when the projected gain across the planning window clears the cost. On a strong squad those high-conviction hits are rare - little on an elite XV is worth a −4 - so Climber barely uses the budget. On a weak squad they’re everywhere, and Climber spends them. From a good squad Climber took about 3 hits a season and the two finished level; from a weak squad it took about 7, and the hit budget is what separates them.

Pick Climber if your starting squad has work to do - which describes most managers’ GW6 squad after a couple of bad early picks - and you are willing to take hits to fix it.

Two earlier presets were retired before the November 2026 data rebuild: a “Balanced” preset, after every differential mechanism we tested lost in all three seasons we then tested on, and a shorter-window (h=4) variant. The current pair was first chosen from a walk-forward sweep of 48 strategy settings and re-checked on the rebuilt data - see the accuracy page for both.

Data sources and refresh cadence

  • FPL official API - prices, injuries and availability flags, refreshed hourly; match stats, ratings and projections, refreshed after each gameweek’s last match and twice a day (10:30 and 23:30 UTC).
  • Past seasons - FPL’s per-match records for every season from 2019/20, from a public archive of past FPL data.
  • Expected goals and assists - Understat for 2019/20 to 2021/22, since FPL published no per-match xG before 2022/23, and FPL’s own figures from 2022/23. Weeks with no published xG count as unknown, not zero.
  • Other competitions - cup, European and other leagues’ matches from FBref, each weighted by the strength of the competition, so a new signing arrives with a rating.
  • Player identity - every source is matched to a player by FPL’s permanent player code, never by name, so two players who share a name never share data and a renamed player stays one player.
  • Ratings and model fits - the overall rating updates after every gameweek. The five component ratings, the blend weights and the difficulty table are refitted by hand during the season, and hold still between refits. The next-week, minutes and captain models are trained on past seasons only.
  • Sitemap & player pages - lastmod stamped from the underlying refresh.

When we test a past season, the model has never seen it: testing 2022/23 uses a model trained on 2019/20 to 2021/22 only, and each later test season adds the one before. The accuracy page has the details.

What this model cannot do

Five things, kept honest:

  1. Chip timing. The planner plans around any chip you set in a week, but it doesn’t choose when to play one. None of our season backtests plays a chip either, so chip handling is checked by tests, not by the validated results.
  2. Mini-league rank. The planner maximises expected points, not your rank in a mini-league. Differentials - low-ownership picks that move you up the rankings when they haul - would need rank as the objective, and we haven’t run that study. Simply boosting low-ownership picks lost points in backtest.
  3. Team news. Expected minutes come from minutes played, starts and who is fit at his position, plus FPL’s flags, not from press conferences. A rotation confirmed two hours before the deadline doesn’t reach the projections until FPL flags it and the projections next refresh.
  4. Set-piece role changes. The captain model reads FPL’s penalty and set-piece order for the next gameweek, but the ratings behind every projection learn about a new taker only from his points.
  5. Short matches and bench weeks. A match under 21 minutes, including a week on the bench, counts as a quarter-weight blank for every component rating except Defender. Backtests showed a gain, but it holds back a player returning from a long absence, or a breakout player who started on the bench, for a few matches.