FPL Predicted Points: How the Model Works and How Accurate It Is

The 2026/27 scorecard, updated every gameweek

Every prediction below was published before its deadline and scored afterwards against what the player actually returned. Across 4 gameweeks and 2,495 player-fixtures, the model's average error is 1.209 points per player per gameweek, with a rank correlation of 0.69. In the 2 gameweeks where FPL published its own expected points for comparison, this model had the lower error in 2 of 2.

Both figures are measured on the same rows, so they are directly comparable — lower is better. In GW4 the model's error was 1.173 against FPL's 1.234.

GW4 predicted points: the model's top 50, and what they actually scored

Its highest-rated pick was Haaland at 6.73 predicted; he returned 9. Its top ten captured 46 of the 135 points available from the ten best-returning players that week (34%). The misses are left in — a table showing only the weeks it got right would not be worth much.

Scored from the snapshot taken before the GW4 deadline, on model v5-serve-honest-2026-08. One gameweek is a single noisy sample — the season figures above are the ones worth judging it on.

What the model predicts

One number per player per fixture: their expected FPL points, for every player in the game, over the next 5 gameweeks. Double gameweeks get a prediction per fixture. The model re-runs each morning and again in the final hours before every deadline, so the run closest to each deadline has final team news, up-to-date injury flags, and the latest bookmaker odds baked in.

You'll see the output in two places: the Predicted Points table (filter by gameweek and position, see the model's captain pick), and inside the Transfer Planner, where every planned squad shows its predicted starting-XI total.

What goes into it

The model is a gradient-boosted ensemble trained on five seasons of historical FPL data - roughly one row per player per fixture, with everything known strictly before kick-off:

• Player form - rolling averages of points, minutes, xG, xA, ICT, BPS and more over recent matches and the season to date.
• Fixtures and team strength - home/away, recent goals for and against for both teams, and FPL's strength ratings.
• Bookmaker odds - win/draw/loss and over-2.5-goals probabilities for each match. Bookies price in news faster than any stats site.
• Injury and availability status - flags, chance-of-playing percentages, and whether there's news attached to a player.
• Crowd signals - ownership levels and transfer momentum, which often front-run team news.
• FPL's own published expected points (the xPts figure in the game) - yes, the thing we benchmark against is also an input. The model learns when to trust it and when to override it.

We deliberately keep the exact feature engineering and model architecture out of this post - that part is our edge. But the data categories above are the complete list of what the model can see.

How we tested it (without cheating)

The easiest way to fake a good prediction model is to test it on data it was trained on. So we did the standard thing properly: the final 19 gameweeks of 2025/26 (GW20–38, 15,572 player-fixtures) were locked away as a held-out test set. All model design, feature selection and tuning happened on earlier data only. The test set was opened at the end to produce the numbers below.

Every prediction in the backtest used only information available before that gameweek's deadline - no hindsight about lineups, injuries, or results.

The results vs FPL's xP

First, a note on terms: throughout this post, xP means the official expected-points figure that FPL itself publishes for every player (the "xPts" you see in the game, ep_next in the API) - not a third-party model built from xG and xA. It's the natural benchmark because it's the prediction every FPL manager already has access to.

Mean absolute error (MAE) is the average number of points a prediction misses by - lower is better.

The captain row is the one that matters most in practice: picking the model's highest-predicted player as captain each gameweek returned 57% more points than picking xP's highest-rated player - and that's before captaincy doubles it.

Where xP still wins (and why)

Look at the last row of that table. FPL's xP is better at ordering nailed starters against each other, even though its point values are less accurate. The reason is simple: FPL sets xP with ground-truth, real-time availability information about every player. Our model reads the same public signals you can - status flags, press conference news, odds moves - and those always lag slightly.

We'd rather publish that than hide it. In practice the two views are complementary: our model tells you how many points a situation is worth far more accurately; xP occasionally knows who's definitely playing a few hours earlier.

How close is this to the ceiling?

Closer than you might think. FPL Review, whose accuracy research we rate highly, estimated what a perfect prediction model could achieve on active players, given that football outcomes are inherently random: roughly 2.81 RMSE and 1.96 MAE. Measured the same way, our starters-only predictions come in around 2.79 RMSE and 1.95 MAE - the same scale. (We quote both metrics deliberately: RMSE punishes big misses more heavily than MAE and always reads higher, so comparing one model's MAE to another's RMSE would flatter it.) To be clear, that doesn't mean we've matched perfection - the two studies cover different seasons and define "starters" slightly differently - but it does mean the remaining gap is small.

Comparing per-outcome error against the two best-documented public models (FPL Review's and the OpenFPL academic model), we also land in the same accuracy class - with the usual caveat that different models were tested on different seasons, so it's a comparison, not a controlled head-to-head.

The takeaway: most of the remaining prediction error isn't model weakness, it's irreducible football randomness. Which leads to the most important section of this post.

How to actually read a prediction

A prediction of 5.2 points does not mean the player will score 5. It means that across many repeats of this exact fixture, he'd average about 5.2 - sometimes 2, sometimes 13. Predictions are averages over possible futures, not fortune-telling.

• Use gaps, not decimals. A 0.3-point difference between two players is noise. A 1.5-point gap is signal.
• Sum across gameweeks for transfers. The 5-GW total view is the right lens for a transfer decision; the single-GW view is the right lens for captaincy.
• Check the freshness stamp. The page shows when the model last ran. Predictions from before a press conference are worth less than ones from after.

FAQs

How accurate are FPL Pulse's predicted points?

On a held-out test of the final 19 gameweeks of 2025/26 (15,572 player-fixtures the model never saw during training), the average prediction error was 0.81 points per player per gameweek, versus 1.05 for the official expected points (xPts) that FPL publishes in the game. For likely starters the gap was bigger: 1.95 points of error versus 2.75. The model's top captain pick each gameweek averaged 7.79 points versus 4.95 for xPts' top pick.

Is the model better than FPL's own xP?

On most measures, yes. By xP we mean the official expected points (xPts) FPL publishes for every player in the game - our model had lower error than it on every calibration metric we tested, and much better captain picks. The honest exception: xPts is still slightly better at ranking nailed starters against each other, because FPL has real-time ground-truth availability data. We publish both results.

What data does the model use?

Five seasons of historical FPL data covering player form, minutes, underlying stats (xG, xA, ICT, BPS), fixtures and team strength, bookmaker match odds, pre-deadline injury and availability status, ownership and transfer trends - and FPL's own xP, which the model uses as one input among many.

How often are predictions updated?

Once every morning as a baseline, plus three more runs inside the final three hours before each deadline. We concentrated the runs near deadlines after measuring that the same model predicts materially better 1.5 days out than 6 days out - late team news is worth more than frequent updates. The Predicted Points page shows exactly when the model last ran.

Can the model predict hauls?

No model can reliably predict a specific 15-point haul - football is too random. What a good model does is put the right players at the top of the list more often. Predictions are best read as averages: a player predicted 5.2 points would average about that across many repeats of the same fixture, with plenty of spread either side.

Is Predicted Points free?

Predicted Points is a Pulse Pro feature (£2.99/month or £24.99/year, 7-day free trial). Pro also unlocks the predictions inside the Transfer Planner, so you can see the model's projected points for every planned squad.