Accuracy
Our predictions are locked before every deadline and checked once the gameweek ends.
The model beats a simple average on both clean sheets and match results. Replayed over last season, its captain picks scored 98 points more than a form pick.
- Clean sheets, better than a simple average
- 5.5%
- Error 0.1809 vs 0.1913 (lower is better)
- Match results, better than a simple average
- 5.4%
- Error 1.0266 vs 1.0852 (lower is better)
- Captain scored, last season
- 246
- Form pick scored 148, before armband doubling
- Matches predicted
- 380
- Replayed week by week, 2025/26
Player predictions, week by week
Every deadline, each model version's predictions are locked before kickoff, then checked against real points when the gameweek finishes. Nothing is recomputed after the fact. The bar is to beat FPL's own predictions and a last-5 form average, week after week.
Ranked players better than FPL: 4 of 4 weeks
Market odds cover 0 of 10 fixtures this gameweek; the rest use our team-strength ratings until bookmakers publish prices, usually on Friday.
| Gameweek 6 | Ranking accuracy (higher is better) | Average miss (pts) | Captain scored |
|---|---|---|---|
| Market fixtures + learned minutes live now | 0.30 | 2.27 | 1 pts |
| Market fixtures | 0.23 | 2.25 | 1 pts |
| FPL's own projection benchmark | 0.17 | 2.97 | 6 pts |
| Learned minutes | 0.17 | 2.32 | 1 pts |
| Learned minutes + team news | 0.17 | 2.32 | 1 pts |
| Team strength + defensive contributions | 0.14 | 2.30 | 1 pts |
| Last-5 form average benchmark | 0.12 | 2.71 | 6 pts |
Earlier gameweeks (GW5, GW4, GW3, GW2)
| Gameweek 5 | Ranking accuracy (higher is better) | Average miss (pts) | Captain scored |
|---|---|---|---|
| Market fixtures + learned minutes live now | 0.20 | 2.41 | 6 pts |
| Learned minutes | 0.18 | 2.41 | 6 pts |
| Learned minutes + team news | 0.18 | 2.42 | 6 pts |
| FPL's own projection benchmark | 0.14 | 2.69 | 5 pts |
| Last-5 form average benchmark | 0.14 | 2.69 | 5 pts |
| Market fixtures | 0.09 | 2.46 | 6 pts |
| Team strength + defensive contributions | 0.07 | 2.45 | 6 pts |
| Gameweek 4 | Ranking accuracy (higher is better) | Average miss (pts) | Captain scored |
|---|---|---|---|
| Learned minutes | 0.30 | 2.67 | 9 pts |
| Market fixtures + learned minutes live now | 0.30 | 2.67 | 9 pts |
| Learned minutes + team news | 0.30 | 2.67 | 9 pts |
| Team strength + defensive contributions | 0.24 | 2.69 | 9 pts |
| Market fixtures | 0.24 | 2.69 | 9 pts |
| FPL's own projection benchmark | 0.16 | 2.79 | 1 pts |
| Last-5 form average benchmark | 0.16 | 2.81 | 1 pts |
| Gameweek 3 | Ranking accuracy (higher is better) | Average miss (pts) | Captain scored |
|---|---|---|---|
| Market fixtures + learned minutes live now | 0.23 | 2.33 | 9 pts |
| Learned minutes | 0.22 | 2.33 | 9 pts |
| Learned minutes + team news | 0.22 | 2.33 | 9 pts |
| Team strength + defensive contributions | 0.14 | 2.44 | 9 pts |
| Market fixtures | 0.12 | 2.45 | 9 pts |
| Last-5 form average benchmark | -0.02 | 3.14 | 2 pts |
| FPL's own projection benchmark | -0.02 | 3.17 | 2 pts |
| Gameweek 2 | Ranking accuracy (higher is better) | Average miss (pts) | Captain scored |
|---|---|---|---|
| Team strength model | 0.46 | 2.47 | 23 pts |
| FPL's own projection benchmark | 0.28 | 2.59 | 6 pts |
| Last-5 form average benchmark | 0.12 | 3.56 | 4 pts |
Where we differ from FPL
Calibration
Points on the diagonal mean predicted probabilities match how often things happened.
Show the calibration charts
Predicted probabilities match what happened.
Data table
| Predicted | Observed | n |
|---|---|---|
| 8% | 14% | 21 |
| 16% | 15% | 210 |
| 25% | 25% | 287 |
| 34% | 31% | 189 |
| 44% | 51% | 45 |
| 53% | 50% | 8 |
Predicted probabilities match what happened.
Data table
| Predicted | Observed | n |
|---|---|---|
| 17% | 13% | 24 |
| 25% | 28% | 54 |
| 35% | 40% | 101 |
| 45% | 37% | 91 |
| 55% | 59% | 61 |
| 64% | 70% | 37 |
| 76% | 75% | 8 |
Replayed over a full season
Every gameweek of last season, replayed: each model version rebuilt from only the information available before that deadline, then checked against real points. FPL's own predictions are not published for past seasons, so that comparison lives in the weekly table above; the last-5 form average is here.
The full model's weekly captain pick outscored the form average's by 98 points over the season, before armband doubling, and ranked players better in 30 of 38 gameweeks.
Show every model version
| Model version | Ranking accuracy (higher is better) | Average miss (pts) | Captain scored, season |
|---|---|---|---|
| Last-5 form benchmark | 0.124 | 2.68 | 148 pts |
| Statistical model | 0.203 | 2.53 | 222 pts |
| With learned minutes | 0.224 | 2.42 | 239 pts |
| With market odds | 0.223 | 2.52 | 220 pts |
| Learned minutes + market odds | 0.240 | 2.41 | 246 pts |
| Plus relative fixture scaling | 0.241 | 2.40 | 243 pts |
| Plus joint calibration (cards, saves, assists) | 0.237 | 2.37 | 243 pts |
Method
How these were tested
Every prediction was made without seeing the result: for each match date of 2025/26, the team strength model is rebuilt from earlier matches only (two prior seasons plus the season to date, recent games counting more) and predicts that day's games. The simple average gives every match the historical average rates. The predictions across the site come from the same model, built on all current data.