How Our Expected-Points Model Works (and Where It Is Weak)
By Geomi George · Published 14 November 2026
Every recommendation on this site, from the Best XI to the transfer ideas on your own team, comes from one expected-points model. This article explains how it works, in enough detail that you could reproduce it, and then sets out where it is weak. We publish it because advice is only worth following if you can see what it is built on and how it has performed.
What the model is trying to do
For each player and each gameweek, the model estimates how many fantasy points he is likely to score. It does not try to predict a single match. Football is too random for that: a penalty, a red card or a deflection can swing a score by ten points. What can be estimated is the average outcome if the same match were played many times. Over a season those averages are what separate good decisions from lucky ones.
The estimate is built in two stages. First the model works out a player's base strength: the points he would be expected to score in an average fixture. Then it adjusts that figure for the actual opponent in each gameweek and for the chance that he will not play.
Stage one: base strength
Base strength blends what a player has already done with what his underlying numbers say he should do next.
| Ingredient | Weight | Why it is there |
|---|---|---|
| Points per game this season | 45% | The broadest evidence of what a player produces |
| Form (points per match, last 30 days) | 20% | Picks up a change of role, system or fitness |
| Bonus points per 90 minutes | 10% | Some players collect bonus far more reliably than others |
| A forward-looking term (below) | 25% | Chances and defensive quality predict better than past goals |
A very small additional term, half a per cent, comes from the game's own ICT index. The forward-looking term depends on position. For midfielders and forwards it is expected goal involvement per 90 minutes, converted into points at the value of a goal or assist for that position. A midfielder creating and receiving good chances is credited for them whether or not they have gone in yet. For defenders and goalkeepers it is the probability of a clean sheet, derived from the expected goals the team concedes per 90 minutes: a side conceding half an expected goal a game keeps a clean sheet about 70% of the time by this measure, a side conceding two about 25%. Goalkeepers get a small extra credit for the saves a busier defence brings.
Midfielders and forwards also receive a fixed addition for set-piece duties: about 0.6 of a point per match for a first-choice penalty taker, 0.15 for a second-choice taker or a first-choice direct free-kick taker, and 0.1 for first-choice corners. Current duties are listed on the set-piece takers page.
Two corrections that stop the model fooling itself
Minutes. Points only arrive if a player is on the pitch. A player averaging 75 minutes or more per gameweek keeps his full rating. Below that it is scaled down in proportion, and a player who is clearly not a regular starter is cut to roughly a third. All per-90 figures are calculated on at least one full match, so a substitute who scored in a ten-minute appearance cannot be rated as the best player in the league.
The start of the season. After one gameweek, points per game is a single afternoon. Until six gameweeks have been played, the model blends its rating with a prior based on a player's price and ownership, which between them express what the game and its managers expect of him. The prior counts for 75% after one gameweek, 45% after three, 15% after five and nothing from the sixth onward. This is why the early Best XI pages carry an "early-season mode" notice.
Stage two: fixtures and availability
Each fixture has an official difficulty rating from 1 to 5. The model multiplies base strength by a factor for the rating:
| Difficulty | 1 | 2 | 3 | 4 | 5 |
|---|---|---|---|---|---|
| Attackers | ×1.5 | ×1.3 | ×1.1 | ×0.9 | ×0.7 |
| Defenders and goalkeepers | ×1.5 | ×1.3 | ×1.1 | ×0.72 | ×0.56 |
Defenders are penalised more heavily in hard fixtures because most of their points come from the clean sheet, and that is what a strong opponent takes away. A double gameweek counts both matches; a blank counts nothing.
A player with an injury flag is discounted by his stated chance of playing. A 75% flag removes a quarter of his projection for the coming gameweek and an eighth for the weeks after, on the reasoning that a knock this week has usually cleared within three. Players rated below 75%, and those who are injured or suspended, are not recommended at all.
From player ratings to a squad
The Best XI turns the ratings into a legal squad: fifteen players, two goalkeepers, five defenders, five midfielders and three forwards, no more than three from one club, within £100m. It fills the squad from the highest-rated players downwards, always holding back enough money to complete the remaining positions, and then spends anything left over on upgrades. Early in the season, when ratings are least reliable, it gives each position a fixed share of the budget instead, so the money cannot all be drawn into attack. From the fifteen it picks the best legal starting eleven, captains the highest-rated starter and orders the bench.
The same ratings drive everything else. My Team projects your own players over the next four gameweeks and states each suggested transfer as a gain in points, so it can be weighed against a four-point hit. The transfer planner compares a multi-week plan with making no transfers. The squad builder picks the best eleven from a draft. Because they share one set of ratings, they point the same way.
Where the model is weak
We would rather you knew these than discovered them.
- It cannot read team news. It knows a player is flagged, not that his manager said on Friday that he will start, or will be rested. A human reading the press conferences will beat it on any player whose situation changed that week.
- It reacts to form; it does not anticipate it. Thirty days of results feed the rating, so the model is late to a breakout and late to a decline.
- Difficulty ratings are blunt. They describe an opponent's general strength, not its current form, and they are the same for attack and defence. A team that concedes freely but scores a lot is a good fixture for your forward and a bad one for your defender; one number cannot say both. Our weekly previews regularly disagree with a rating for this reason.
- It does not model rotation. A player who starts three matches in four looks like one who is taken off early. During congested periods it overrates the first and underrates the second.
- One-week predictions are mostly noise. The model's edge, where it has one, shows over runs of fixtures. Its captain pick for a single gameweek will be wrong often, as anyone's will.
- It ignores ownership. The model picks the highest expected score. It does not know that a blank from a player owned by 80% of managers costs you nothing against your rivals, while a blank from a 5% pick costs you everything. We add that judgement by hand in the previews.
How it has actually done
The backtest records each week's Best XI before the deadline and scores it afterwards. Over the first ten gameweeks of this season it scored 564 points against an average manager's 563: level, in other words. Its best week was Gameweek 10, twenty points above the average; its worst was Gameweek 2, twelve below. That is an honest description of a model that rates players sensibly and has no magic in it, and it is a test of the ratings only, since the Best XI is rebuilt each week with a full budget in a way no real manager could follow.
We treat the model as a well-informed starting point and say so when we overrule it. After Gameweek 4 we adopted a rule that when two captain options are within two projected points, the more widely owned one is the better bet. In Gameweek 7 that rule put us on Haaland when the model preferred another player, and it was right. In Gameweek 8 we overruled the model's decision to bench two defenders for a hard away match, and the model was right. Both are recorded in the gameweek reviews, which mark every preview call after the event.
Using it well
Trust the model most on questions it is built for: which of two similar players has the better underlying numbers and fixtures over the next month. Trust it least on anything that turned on news from the last forty-eight hours. Use a horizon of four to six gameweeks for transfers and one gameweek only for your captain and bench. And check its work: the numbers it uses are all visible on the players, fixtures and xG regression pages, and the concepts behind them are explained in Understanding xG and xGI and How Fixture Difficulty Really Works.