How the forecasts are made
Model 1.4.5. Trained on completed Gameweeks; FPL data fetched —. Historical evidence is not loaded, so confidence is lower.
What goes into a player's expected points
- Playing time first. Each fixture is modelled as a start, a substitute appearance or no appearance, using starts, minutes and official availability.
- Scoring by position. Goals, assists, clean sheets, saves, defensive contributions, bonus and cards are scored under FPL rules for that position.
- Current form, steadied by history. This season's underlying numbers (xG, xA) are blended with previous seasons. This season counts for more as minutes build up.
- Fixture context. Opponent strength, home or away, and fixture difficulty adjust each match.
- Double and blank Gameweeks. A Gameweek with two matches counts both; a blank scores zero.
Position calibration
Raw points are adjusted with a separate straight-line calibration for each position. The adjustment was fit using walk-forward forecasts from 2024-25 and then tested on 2025-26 without refitting. It corrects persistent over- or under-prediction; it does not use a player's actual points from the Gameweek being forecast.
| Position | Adjustment | Training player-weeks |
|---|---|---|
| GKP | -0.58 + 0.929 × raw xPts | 2,654 |
| DEF | -0.31 + 0.859 × raw xPts | 8,470 |
| MID | -0.15 + 0.962 × raw xPts | 11,644 |
| FWD | -0.32 + 1.045 × raw xPts | 2,788 |
On the 27,102 player-Gameweeks in the held-out season, all-player MAE changed from 1.232 to 1.049. Among players who appeared, within two points rose from 67.8% to 70.9%, but RMSE worsened from 2.983 to 3.102. A flat two-point forecast reached 77.2% within two. The calibration training run used an earlier double-Gameweek scoring method and needs refitting before these gains can be relied on.
Playing-time evidence
When a player has no minutes after their team has played at least two matches, and no historical starting record is available, the model gives less weight to its pre-season playing-time assumption. Players who have appeared or have a known previous-season role keep the existing estimate. In a reconstructed 2025-26 replay, this reduced all-player average error from 1.163 to 1.051 points and appearance-probability error from 0.153 to 0.126. Accuracy within two points among players who appeared stayed near 70.7%. The available 2026-27 GW4-6 replay also improved all-player error, though that sample is small.
Goalkeeper saves
Save points now count complete groups of three saves per match, as FPL awards them. The model estimates the chance of reaching each three-save threshold, including the chance that the goalkeeper does not play. Historical checks improved average error slightly but raised the size of some goalkeeper misses, so this remains under review.
Expected points, range and median
Expected points is the average outcome, not the most likely exact score. The likely range runs from the 10th to the 90th percentile of simulated outcomes, so roughly 1 in 10 results fall below it and 1 in 10 above. The median is the middle outcome and is often lower than the average because big hauls pull the average up.
Risk, confidence and data quality are different
- Risk describes how wide the range is for that player and horizon.
- Confidence describes how strong the evidence behind the forecast is. It is not the chance of being right.
- Data quality is a 0–100 coverage score for the inputs. It is not an accuracy percentage.
From forecasts to decisions
- Transfers compare projected points over five Gameweeks, subtract 4 points per transfer beyond your free ones, and keep budget, positions and the three-per-club rule legal. The search covers the top candidates per position, not every possible squad.
- Holding is a valid answer. If the best move doesn't beat the hit, FPL Prism says so.
- The Planner compares rolling with one- and two-transfer moves across eight Gameweeks.
- Chip guidance flags plausible windows; it does not calculate the points a chip would add.
Betting-market prior
An optional prior from betting markets can nudge team-level goal and clean-sheet expectations. It is not applied in the current forecasts.
No betting-market observations are currently connected.
Limitations
Team news, rotation, injuries and tactics change after forecasts are made. Players with few minutes have thin evidence. Public data doesn't include your selling prices or remaining free transfers. Treat every number as an estimate.
Track record
Forecasts update until each fixture starts. The separate deadline snapshot is preserved for independent comparisons in the Modelbook.