How accurate are the predictions?
Every prediction on this site is written down before the deadline and never recalculated. This page is what happened next: what the model said, what the players actually scored, and how far apart the two were.
The short version
Across 105 settled player-gameweeks, the projections were out by 2.76 points on average. That is 9.9% better than the best you could do knowing nothing about the players — predicting the same score for everybody, which would be out by 3.06. Ranked against each other rather than scored absolutely, the correlation between projection and outcome is 0.568.
What is being measured
Each row below is one squad, and each squad contributes fifteen predictions a gameweek — one per player. A prediction counts only once its gameweek is settled and Fantasy Premier League has confirmed the bonus points, so nothing here is provisional.
| Sample | Predictions | Avg error | No-model baseline | Better by | Bias | Rank corr. |
|---|---|---|---|---|---|---|
|
AI Best XI
the squad the optimiser picks fresh each week, so the players the model rates highest
|
105 | 2.76 | 3.06 | 9.9% | +0.53 | 0.568 |
|
AI Manager
one squad carried across the season, so it holds players bought several weeks earlier
|
105 | 2.94 | 2.91 | -0.9% | +0.31 | 0.174 |
|
Real managers
squads captured at the deadline from people using the site - picked by humans, so the least flattering sample here
|
225 | 4.0 | 4.03 | 0.8% | -1.97 | 0.206 |
Avg error is the mean absolute difference between predicted and actual points. Bias is the average signed difference: a positive number means the model projects more than players deliver. Rank correlation is Spearman's — whether the model puts players in the right order, which is the claim this site actually makes. A figure marked * has fewer than 60 predictions behind it.
Gameweek by gameweek
One averaged number hides the shape of the thing. The model has good weeks and bad ones, and the spread is more use than the average when you are deciding how much to trust a projection.
| Gameweek | Predictions | Avg error | Bias | Avg predicted | Avg actual |
|---|---|---|---|---|---|
| GW1 | 15 | 2.35 | +1.41 | 3.88 | 2.47 |
| GW2 | 15 | 3.93 | -0.2 | 4.0 | 4.2 |
| GW3 | 15 | 2.45 | +2.06 | 4.13 | 2.07 |
| GW4 | 15 | 3.65 | -1.15 | 3.65 | 4.8 |
| GW5 | 15 | 2.19 | +0.98 | 3.78 | 2.8 |
| GW6 | 15 | 2.35 | +0.61 | 3.47 | 2.87 |
| GW7 | 15 | 2.36 | -0.03 | 3.51 | 3.53 |
The squads themselves
A different question, and worth keeping separate: the table above asks whether the projections are any good, this one asks whether the squad built out of them worked. A squad can score well in a week the individual projections were poor, and often does — it only needs the order to be roughly right, not the numbers.
| Gameweek | Projected | Actual | Difference |
|---|---|---|---|
| GW1 | 60.2 | 36 | -24.2 |
| GW2 | 61.4 | 64 | +2.6 |
| GW3 | 62.5 | 34 | -28.5 |
| GW4 | 54.8 | 81 | +26.2 |
| GW5 | 57.5 | 48 | -9.5 |
| GW6 | 52.9 | 49 | -3.9 |
| GW7 | 52.2 | 65 | +12.8 |
How to read all this
The sample is picked, not random. These are squads somebody chose, and mostly squads the model chose for itself. Scoring well on players it rated highly is easier than scoring well on the whole pool, because most of the pool is players who never get near a pitch and are trivially projected at nearly nothing. So read these figures as how good are the projections for players worth considering — which is the useful question, but is not the same as a score over every player in the game.
The predictions are risk-adjusted. What gets stored against a pick is the chance the player features multiplied by what he would score if he did, so a doubtful player who sat out and scored nothing was partly anticipated. That is the fair comparison to make, and it is also why these numbers sit a little below the raw projections on a player's own page.
An average error of two points is not a small number when most players score between two and six. That is the honest state of predicting football over ninety minutes from a handful of inputs, and it is why the FAQ says the projections answer "who is the better pick" rather than "how many will he get". The rank correlation is the figure that speaks to the first question; the average error speaks to the second, and it is the weaker of the two.
Nothing on this page is recomputed when you load it. The predictions were frozen at their deadline and the actuals were written once the round settled, which is what makes the comparison worth publishing at all — see how the model works for why that freeze matters.