In short
Gaffer Lab rates all 20 Premier League clubs with a statistical goals model built on four seasons of results and expected goals (xG), projects every FPL player's points fixture by fixture from his minutes and his Opta underlying numbers, plays each gameweek 400 times in a Monte Carlo simulation to show the risk around that projection, and searches tens of thousands of legal squads to find the strongest 15. The club model's match forecasts have been tested on 800 matches and score within 0.001 log-loss of the bookmakers' closing odds. Nothing on the site is hand-picked.
The data
Every figure on Gaffer Lab starts as a real, published statistic. We use three sources and we do not alter any of them.
| Source | What it provides | How it is used |
|---|---|---|
| Official Fantasy Premier League feed | Prices, ownership, points, minutes, starts, bonus, injury news and chance of playing, fixtures, official difficulty ratings, price-change progress, transfer counts, and any public team's picks by Team ID. | The live state of the game. On the hosted site it is refreshed every six hours; the time of the data is shown on the Overview. |
| Opta match stats, as published through FPL | Expected goals (xG), expected assists (xA), expected goals conceded (xGC) and defensive contributions, per player and per gameweek. Opta is the Premier League's official data provider. | Shown untouched, so they match the official site, and used as the evidence for each player's scoring rates. |
| Match history, four seasons | Results and match-level expected goals for every Premier League match since 2023-24, and each player's previous seasons, from the open Fantasy-Premier-League dataset and football-data.co.uk. | Fits the club ratings and steadies early-season player projections. |
| Bookmakers' closing odds | The final pre-match prices for every match in the history. | Testing only. Odds are the yardstick the model is measured against; they are never an input. |
The engine, in four parts
The engine is called Gaffer AI. It is not a chatbot and it does not guess: it is four statistical models that run in order, each feeding the next, inside your browser, every time the data loads.
Part oneFormbook: how strong is every club?
Formbook gives each club an attack rating and a defence rating, where 1.00 is a league-average side. It is a time-weighted Poisson goals model in the Dixon-Coles family, the same class of model long used to price football matches. Each match counts as a blend of the expected goals created and the goals actually scored, because chances created predict future scoring better than the scoreline does over short spans. Recent matches weigh more than old ones, home advantage is fitted from the data, and promoted clubs start from FPL's own rating until their results arrive.
From two clubs' ratings and the venue, Formbook produces the probability of every scoreline. Clean-sheet chances, the odds of scoring twice or more, projected goals and Lab FDR, our 1 to 5 fixture difficulty rating, are all read from that one grid.
Part twoPlaymaker: how many points will he score?
Playmaker turns each player into an expected-points figure, fixture by fixture:
- Minutes first. The chance of starting comes from the player's recent starts, with the last few matches counting most, scaled by FPL's injury flag. Suspensions, dated returns from injury and the risk of a yellow-card ban are applied to the specific gameweeks they affect.
- Rates second. Goal and assist rates come from his Opta xG and xA, adjusted for the strength of the defences he has faced. Early in a season a few lucky matches can distort any rate, so each one is blended with a sensible starting estimate from the player's own previous seasons and his price, and the real evidence takes over as the minutes build. Statisticians call this empirical-Bayes shrinkage.
- Fixture third. Every rate is scaled by Formbook's view of the opponent and the venue.
Part threeReplay: how wide could the week be?
Two players on 6.0 expected points can be very different bets. Replay runs a Monte Carlo simulation: it plays each player's next gameweek 400 times, drawing whether he starts, his minutes, goals, assists, clean sheet, goals conceded, bonus and cards from their probabilities. The spread of those 400 scores gives the Floor (a bad week, the 10th percentile), the Ceiling (a good week, the 90th percentile), the Range between them and Haul%, the chance of 10 points or more. A squad-level version plays a whole team's week, with team-mates' clean sheets tied together as they are in real matches.
Part fourScout: what is the best squad?
Scout is a mathematical optimiser. Inside the £100m budget, the 2-5-5-3 squad shape and the three-per-club limit, it builds a squad, then improves it swap by swap, restarts from several different starting points, and keeps the best. Around 50,000 legal squads are scored per build, and the same search ranks transfers for a loaded team by how much each move lifts the best eleven, points hits included. Only players likely to be available and to start are considered.
How expected points add up
Expected points (xP) are the FPL scoring rules applied to probabilities. For each fixture:
| Part of the score | What drives it |
|---|---|
| Appearance | The chance he plays at all, and the chance he plays 60 minutes. |
| Goals and assists | His opponent-adjusted xG and xA per 90, his expected minutes and the fixture, times the points his position earns. |
| Clean sheet | Formbook's clean-sheet probability for that match, times the chance he plays 60 minutes. |
| Goals conceded and saves | The goals his club is expected to concede, and for goalkeepers the saves that come with them, worked out exactly rather than rounded. |
| Defensive contributions | The chance he reaches the threshold for the +2 (10 for defenders, 12 for midfielders and forwards), from his own record and the opponent. |
| Bonus and cards | His bonus and yellow-card rates per 90. |
The parts are added together and summed over the next one, five or eight gameweeks. Next gameweek's figure also leans a little on FPL's own estimate, which reacts quickly to very recent form. The same stored number is read by every page, so the simulator, My Team and the tables cannot disagree.
How we test it
A model that is never checked is an opinion with decimals. Ours is backtested: it is made to forecast past matches using only what was known before each kick-off, and scored on what then happened.
Club model against the bookmakers
800 Premier League matches from August 2024 onwards, with the model refitted before every match date. Bookmakers' closing odds are the sharpest public forecast in football, so they are the benchmark. Lower is better on every measure.
| Forecaster | Log-loss, home / draw / away | Ranked probability score |
|---|---|---|
| Gaffer Lab (Formbook) | 0.994 | 0.201 |
| Bookmakers' closing odds | 0.993 | 0.200 |
| Constant guess (league averages) | 1.084 | n/a |
Clean-sheet chances against what happened
When the model says 30%, about 30% should come in. Across the same 800 matches (1,600 team performances):
| Model said | Average forecast | Clean sheets kept | Fixtures |
|---|---|---|---|
| Under 10% | 7.5% | 7.8% | 116 |
| 10 to 20% | 15.7% | 15.8% | 506 |
| 20 to 30% | 24.8% | 26.5% | 559 |
| 30 to 40% | 34.1% | 35.4% | 297 |
| 40 to 50% | 43.5% | 40.2% | 97 |
Player models
- Who starts. Tested on 17,000 player-gameweeks. Weighting the last few matches most scored a log-loss of 0.452, against 0.517 for a player's season-average start rate.
- Scoring rates. Tested on 32,000 player-gameweeks to set how much early-season evidence to trust. The test also rejected a popular idea: weighting a player's most recent matches more heavily made xG-based predictions worse, not better, so we do not do it.
- Defensive contributions. Measured on a full season: defenders reach the +2 in 30% of away starts against 24.5% at home, so the venue is priced in.
Settings are chosen by these tests, not by feel, and the tests can be re-run whenever new data arrives.
What it cannot see
Being straight about the gaps is part of being trustworthy. The model does not account for:
- Own goals, penalty misses and red cards.
- A manager's surprise rotation, or an injury that has not been announced.
- Injuries with no return date, beyond FPL's own flag.
- A team-mate's absence changing a player's role.
Early in a season the sample is small, which is why history carries more weight in August than in March. And football is a low-scoring, high-luck sport: a projection is the average of how a week could go, not a promise. Use the Floor and Ceiling to judge the risk, not just the xP.
Glossary
- xP (expected points)
- A player's projected FPL score: the average of how his gameweek could go, from minutes, xG, xA, clean-sheet odds, bonus, saves, defensive contributions and cards, adjusted for the fixture.
- xG (expected goals)
- The quality of the chances a player or team had, measured by Opta as the probability each shot is scored. A better guide to future goals than goals themselves.
- xA (expected assists) and xGI
- xA is the quality of the chances a player created. xGI, expected goal involvements, is xG plus xA.
- xGC (expected goals conceded)
- The quality of chances the opposition had while the player was on the pitch. Lower is better for clean sheets.
- DC (defensive contributions)
- Clearances, blocks, interceptions and tackles, plus ball recoveries for midfielders and forwards. FPL awards 2 points for 10 in a match (defenders) or 12 (midfielders and forwards).
- FDR and Lab FDR
- Fixture Difficulty Rating, 1 (easiest) to 5 (hardest). Official FDR is set by FPL. Lab FDR is Gaffer Lab's own, from the opponent's modelled attack and defence, recent form, the official rating and the venue.
- Haul%, Ceiling, Floor and Range
- From the simulation: the chance of 10 or more points, a good week (one in ten goes at least this well), a bad week (one in ten goes this badly or worse), and the span between them.
- EO (effective ownership)
- The share of managers who start a player, counting captains twice. It is what actually moves your rank.
- Gaffer Score
- A rating out of 100 for any squad, measured against the best eleven the model could pick: 85+ Elite, 70 Strong, 55 Solid, 40 Patchy, lower is a Rebuild.
- Differential and template
- A differential is a player few managers own. The template is the squad most managers hold. Gaining rank means being right where the template is wrong.
- DGW and BGW
- A double gameweek is one where a club plays twice; a blank gameweek is one where it has no fixture.
Questions
What is Gaffer Lab?
An independent statistical modelling tool for Fantasy Premier League (FPL). It turns official FPL data and Opta match stats into an expected-points projection for every player, a floor and ceiling from simulation, a difficulty rating for every fixture and a score out of 100 for any squad.
Who is Gaffer Lab for?
Three kinds of Fantasy Premier League manager. Beginners who want easy picks backed by stats, without needing to know what xG means. Experienced and elite managers who want more detail than the official site gives: Opta underlying numbers, floors and ceilings, effective ownership among the top managers. And managers whose rank is sliding, who want to turn it around with informed transfers and captaincy instead of guesswork. All of it sits in one place: stats, clubs, players, your own team and the live gameweek. The guide shows where each of them should start.
Does Gaffer Lab use Opta data?
Yes, as published. The xG, xA, xGC and defensive contribution figures in Fantasy Premier League are collected by Opta, the Premier League's official data provider. We show them untouched and build our own projections on top. Gaffer Lab is not affiliated with or endorsed by Opta or Stats Perform.
How accurate is it?
On 800 past matches the club model scored a log-loss of 0.994 against 0.993 for the bookmakers' closing odds, and its clean-sheet chances matched what happened to within a few points. Individual player scores are far noisier than match results, which is why every projection comes with a floor and a ceiling.
Are the picks chosen by people?
No. Every ranking, signal and recommendation is calculated from the data each time it loads. Nothing is hand-picked, tipped or sponsored.
Is it affiliated with the Premier League or Fantasy Premier League?
No. Gaffer Lab is an independent project, not affiliated with, endorsed by or sponsored by the Premier League, Fantasy Premier League, Opta or Stats Perform. Those names are used only to describe the game the tool is for and the data it uses.
Does it give betting advice?
No. Bookmakers' odds are used only as a yardstick for testing. Gaffer Lab is a tool for picking a fantasy football team.
How do I get in?
Gaffer Lab is invite only for now. If someone told you about it, ask them for their invite link: it lets you straight in. If the front page offers "Request an invite", you can ask there too.
Your mates guess. You'll know.
Bragging rights, backed by numbers. Load any FPL Team ID and see the projection, the risks and the moves. New here? See what every page does, with screenshots.
Open Gaffer Lab