Prediction Accuracy
After a round finishes, we score the predictions against actual fantasy points (FIFA feed) — error, ranking, scouting-bonus calibration, and how the recommended XI did. The hints feed back into tuning the model.
attacking-output multiplier the tuner learned per position (1.0 = no change)
predicted chance to clear threshold vs what actually happened
| Bucket | Predicted | Observed | n |
|---|---|---|---|
| 0-20% | 10% | 12% | 542 |
| 20-40% | 25% | 42% | 156 |
- XI actual points (capt ×2)
- 35.0
- Captain
- Marcus Thuram (1 pts)
- Captain rank in XI
- #8 of 11
- Best captain (hindsight)
- Lucas Digne
- Avg player (field)
- 3.16 pts
where the model flatters players — chasing differentials is only right if the low-owned buckets aren't inflated
- →Under-predicting by 1.05 pts on average — scale up.
- →GK: under-predicting by 2.47 — adjust GK rates.
- →DEF: under-predicting by 1.53 — adjust DEF rates.
Every squad's real score vs the hindsight-best XIs, plus every big miss classified to the model factor that caused it — so each round makes the next projection better.
- FWDJonathan David2.9 →24
Model gave him only 48% to start, but he played and returned 24. The appearance model missed his role.
- FWDCody Gakpo2.8 →19
Model gave him only 54% to start, but he played and returned 19. The appearance model missed his role.
- FWDAyase Ueda2.5 →18
Model gave him only 48% to start, but he played and returned 18. The appearance model missed his role.
- GKEloy Room0.6 →14
Model gave him only 46% to start, but he played and returned 14. The appearance model missed his role.
- FWDIsmaïla Sarr2 →15
Model gave him only 47% to start, but he played and returned 15. The appearance model missed his role.
- MIDJohan Manzambi2.4 →15
Model gave him only 46% to start, but he played and returned 15. The appearance model missed his role.
- DEFNathaniel Brown6.8 →2
Ceiling pick projected 6.8 but returned 2. High-variance blank; the miss cost is amplified if captained.
- DEFNico Schlotterbeck5.6 →1
Played but blanked (1 vs 5.6 projected) — mostly variance.
- FWDMarcus Thuram5.2 →1
Played but blanked (1 vs 5.2 projected) — mostly variance.
- DEFLadislav Krejcí4.8 →1
Played but blanked (1 vs 4.8 projected) — mostly variance.
- DEFJoshua Kimmich5.7 →2
Played but blanked (2 vs 5.7 projected) — mostly variance.
- FWDHarry Kane5.6 →2
Played but blanked (2 vs 5.6 projected) — mostly variance.
- 10×Start-probability model (appearances / rotation signal)
- 9×Variance blank — played, no returns; only act if it repeats
- 1×Ceiling pick blanked — captaincy/exposure sizing, not projection
backtested 4 Aug, 08:15 UTC