The Projected 2026-27 figure on a player page is not a model that beats the field, and we don’t claim it is. We built the deeper version — prior points, minutes, xG/xA, position, team strength — and measured it against simply using last season’s raw total. It tied: 74.3% against 74.4% pairwise accuracy. So we did not adopt it, and the projection still starts from the number you can check yourself — what the player actually did — rather than from a model you cannot.
So for a returning player the projection starts from last season’s actual total — and then makes three corrections to it, each tested on seasons it never saw:
What the projection adds beyond that is the two things last season’s number cannot do on its own:
Every attempt to reorder that list has failed — nine of them now, and we would rather publish them than imply we never tried. Measured the same way each time — pairwise ordering accuracy within a position, on seasons the method never saw:
So within a position, our ranking is still last season’s points ranking. We say that plainly rather than dress it up: the numbers beside it are better than last season’s raw ones, but the ORDER is not, and nothing we have tried has beaten it. What it does add is coverage and disagreement: FPL’s rank is the only number a player new to the league has, and ours gives them one too — from a curve validated on seasons it never saw. And where our projection and FPL’s rank part company we show both numbers and the gap, rather than quietly picking a side. We have now tested our ranking against FPL’s directly, on the one season with a published draft rank: neither beat the other — the difference was well inside the margin of error — so we don’t claim our list is better than theirs. What we can say is that the reverse isn’t true either, which is what the disagreement tabs rest on.
The projection you see combines two things at equal weight: what a player actually did last season, and what FPL’s published preseason rank says about him. We resisted this for a long time, because on our own measure — putting players in the right order within a position — FPL’s rank is not better than last season’s points. We tested it directly and it came out a tie, well inside the margin of error, in both directions.
What changed our mind was a different question. We simulated 800 drafts against three kinds of opponent and asked which basis produced the better squad. The combination won all three — against opponents drafting off FPL’s board, off last season’s points, and with hindsight — while each input alone did not. We also checked it against a control: the same board with random noise instead of real information, tuned to differ from FPL’s by the same amount. The noise version lost badly, so the gain is not simply “be different from the room”.
There is a mechanism behind it, and it is measurable independently. FPL’s rank systematically over-rates players with no Premier League record — its top 150 held 38 of them averaging 60 points, against our blend’s 18 averaging 80 — and that is the same effect our three-season signing discount measures from the other side. The blend keeps FPL’s ordering everywhere else.
The honest limits. The ranking evidence is a tie, not a win — we are not claiming our list is better than FPL’s, or theirs than ours. The draft evidence is one season, because 2025-26 is the only completed season with a published draft rank to test against. We will re-run it when 2026-27 finishes. The weight is a plain 50/50, chosen rather than tuned: the tests showed a wide flat range where any middling weight beat either extreme, and the best-looking weight differed between tests, so tuning it would be fitting to one season’s noise.
A player page now shows three forward-looking numbers, and they come from different places over different lengths of time. The season figure is ours: last season’s actual total, shrunk toward the positional average and blended with FPL’s preseason rank, for the whole season. The next-four-gameweek and next-gameweek figures beside it are also ours, from the separate model measured on this page. The one labelled FPL’s own estimate is not ours at all — it is the number FPL publishes for the single upcoming gameweek, and it is what our waiver engine ranks players on when it tells you who to pick up.
They cannot be reconciled by multiplication. A one-gameweek number times the gameweeks remaining is not the season number, and the difference between the two is not an error in either of them. They answer different questions: one is “what will he do this week”, the other is “what will he do this year”, and they are built from different evidence by different people. If they disagree, that is information about how much is uncertain — not a sum that failed to add up. We show both rather than picking one, and we label which is which, because collapsing them into a single figure would hide exactly that.
Every projection is shown with a typical range, because a single number implies a precision this method does not have. Measured by holding seasons out of the fitting and projecting them blind — 1,346 player-seasons — the average miss is:
That range is a typical miss, not a guarantee. About 6 in 10 players land inside it; roughly 9 in 10 land within twice it. It is not a 95% confidence interval and we do not present it as one.
Nine failed attempts point at something: the next useful thing for a drafter probably isn’t in the box score at all. A player changed clubs. His club got better or worse. The man who supplied his chances left. So beside the projection we show context — facts about what changed around a player, each with its source:
These do not move the number, and that is deliberate. We have no validated size for any of these effects — nine tests say we cannot get one from history — so inventing a multiplier would be exactly the kind of false precision this page exists to avoid. They are facts, stated with their source, that a drafter can weigh and a model can’t. The last one is explicitly a proxy: we hold each club’s assist totals, but not who assisted whom, so “the top creator left” is the honest substitute for a passer-to-scorer link, not the link itself.
We also checked whether these should count for more at any one position — forwards were the obvious candidate. Over six seasons, club movers beat stayers by +3.7 points on average, and the split by position (defenders +6.5, forwards +4.6, midfielders +2.9, goalkeepers −3.3) gave us no reason to single forwards out. So every position gets the same chips on the same terms.
And what it still does not claim: it is not a model that beats the field. A deeper model tied the raw total at 74.3% against 74.4%, and the adjustments above are corrections to a number, not a claim to have out-predicted anyone. Three limits we would rather state than have you discover. The newcomer discount is provisional — highly-drafted arrivals returned about two-thirds of their slot in each of the last three seasons, but that rests on just 42 players, so we apply a gentler cut than we measured. The curve’s scale rests on a single season, because defensive-contribution points lifted scoring ~15% in 2025-26 and only that season plays by the current rules. And availability is shown, never applied: we will tell you a player managed 20 of 38 games and what a full season at that rate would have been, but we do not quietly inflate the projection, because assuming he stays fit is a guess and this page is about not guessing.
Our projection's job is to rank players by the points they'll score. So we measure it the honest way: take the projection, compare it to the points players actually scored, and report the rank-correlation. Higher is better; 0 is noise; 1 is perfect.
The 4-gameweek projection correlates with real outcomes at r = 0.35 (Pearson). The single-gameweek projection at 0.23 (Spearman, among players who actually featured). A single week of football is genuinely noisy, which is why the 4-week figure is the more dependable read — and why our player pages lead with it.
Both numbers come from a leak-free walk-forward test: the model is trained only on earlier gameweeks, then scored on later gameweeks (the back third of the 2025-26 season, ~3,000 held-out player-gameweeks) that it never saw during training. No peeking at the answers.
An earlier version of this model reported r = 0.43. That figure was in-sample — the model was graded on the same data it was trained on, which flatters any model. When we re-tested it properly on held-out gameweeks, the real, out-of-sample number was 0.35. We publish 0.35, not 0.43, because 0.35 is the one that's actually true.
The model is also deliberately conservative: it pulls projections toward the average for each position rather than making bold individual calls. It ranks players better than it spreads them apart — good for ordering who's likely to do well, modest at predicting an exact score. We'd rather under-promise on precision than overstate certainty.
We put the custom model up against the dumbest baseline we could — a player's season-to-date points per game — on the decision the tool actually makes: given two waiver-caliber players at the same position, which one scores more over the next four gameweeks. Out-of-sample, the model won by +1.2 percentage points, with a confidence interval that spans zero — short of the bar we set before running the test. So we are not wiring it into the waiver recommendations; those still use FPL's own published projection.
The nuance, stated honestly: the model does carry more signal than the baseline — it ranks four-gameweek outcomes at 0.264 vs the baseline's 0.207. But that's better correlation, not better calls. Ranking a whole field and picking the better of two close players are different jobs, and on the second one — the one that decides a swap — the extra signal didn't pay off. A model has to earn its place in a recommendation, and on this test it didn't.
What this does not settle: whether anything beats FPL's own projection. That number was never archived historically, so it couldn't be in this test at all. We started capturing it in 2026, and the real comparison runs across the 2026-27 season — see the next point.
Four position-specific models — one each for goalkeepers, defenders, midfielders, and forwards, because what predicts a defender's points isn't what predicts a forward's. Each takes eight inputs from public FPL data: recent expected goals and assists (xG, xA), minutes and rotation, recent form and bonus-point pace, and rolling team strength. It outputs a projected points total.
The data is all public and reproducible — three Premier League seasons of gameweek-by-gameweek FPL Draft stats. No paid feeds, no proprietary sources you can't reach yourself.
The discipline: validation criteria are written down and locked before results are seen, and changes ship only if they clear a pre-set bar. Several proposed improvements were tested this cycle and rejected for not clearing theirs — that's the process that keeps the number on this page honest.
Browse any player's season-by-season record and projection, or build your league.
All players →