← Back to Blog

Model evaluation • GW7

GW7 Predictions vs Actuals: Vušković Delivers Again — and the Model Takes Its Lumps

Illuminated football stadium at night with an empty pitch
Stadium image: Siarhei Nester / Pexels.

Fantasy Matchup Edge has a proper home on the web now. That also means our weekly prediction reviews can sit alongside the FPL projection boards and Sorare rankings they are testing.

Gameweek 7 offered a bit of everything: a repeat success for an unconventional defensive pick, an excellent return from the model's highest-rated FPL player, an uncomfortable miss at the top of the Sorare board, and plenty of goals and fantasy hauls in places the model didn't particularly expect.

Rather than cherry-pick the winners, here's how the locked, pre-kick-off GW7 predictions compared with the actual results — on the same terms we use in our regular model reviews.

The headline: good scoring levels, but plenty of individual noise

Metric FPL Sorare
Clean evaluation sample 200 202
Average predicted score 3.69 50.42
Average actual score 3.87 50.96
Actual minus predicted +0.18 +0.55
Mean absolute error 2.26 13.45
Spearman rank correlation 0.311 0.303

Those average scores were reasonably close: GW7's eligible FPL players scored 0.18 points more than forecast, while Sorare players scored 0.55 more. That is encouraging for the model's overall scoring environment — especially because the numbers are derived from projected football actions and simulated matches, not simply rolled-forward fantasy averages.

But look at the absolute errors and rank correlations too. They show the limits of individual predictions. A good slate-level average does not mean the model knew exactly which players would score the week's biggest hauls.

How we evaluated it: The frozen GW7 archive contained 544 FPL and 588 Sorare player predictions. The comparisons above use players who started, played at least 60 minutes, and had a recorded score on that platform: 200 FPL and 202 Sorare. The other records aren't treated as successful predictions or counted as zeroes. FME projections are explicitly IF PLAYER STARTS, not start-probability or generic-minutes forecasts. This is a performance evaluation of those conditional projections, not a measure of total squad-selection or availability accuracy.

Did the model's highest-ranked players actually outperform?

Yes, although that is only half the story.

Locked pre-GW7 rank Players eligible Their average actual Other eligible players
FPL top 25 12 5.50 3.77 (188 players)
Sorare top 25 12 63.52 50.17 (190 players)

These are genuine pre-Gameweek top-25 ranks, not the top 25 chosen retrospectively from players who happened to play. Only 12 of each platform's ranked top 25 met the starter/60-minute evaluation rule; the other 13 cannot be silently replaced with different players.

Among those who did qualify, the gap was meaningful. The highest-rated FPL selections averaged 5.50 points against 3.77 for the rest, and the top Sorare group averaged 63.52 against 50.17. Useful ranking signal, yes. A guarantee of hauls, absolutely not.

FPL haul probabilities: fairly close in aggregate

Outcome Expected from GW7 probabilities Actually happened
10+ FPL points 12.65 11
15+ FPL points 1.36 1

Those expected counts come from summing each eligible player's pre-match probability, rather than counting how many players the model predicted would definitely hit the threshold.

The 10+ outcome was slightly less common than expected, but not dramatically so. The 15+ comparison was close too, although with only one event we should be cautious about treating it as a strong calibration test.

Sorare probabilities: the 60+ tail ran hotter

Outcome Expected from GW7 probabilities Actually happened
60+ Sorare score 52.28 66
70+ Sorare score 25.69 31
80+ Sorare score 11.72 10

This is the less flattering detail that matters. There were noticeably more Sorare 60+ and 70+ results than the model's probabilities suggested, even though its average projected score was close to the average actual score. At 80+, the model expected slightly more results than appeared.

That is exactly why we publish different tests. A correct mean does not prove that the whole scoring distribution is calibrated. The Sorare threshold curve deserves monitoring over more Gameweeks, not a knee-jerk adjustment based on one weekend.

Vušković again: not just a lucky 11-point return

The standout story for me is Luka Vušković. Against Crystal Palace, the Brighton defender returned 11 FPL points and 77.04 Sorare points. But what makes this especially interesting is what the model said before the match.

Pre-match GW7 signal FME model
FPL overall rank 9th
FPL expected points 5.50
Chance of 10+ FPL points 16.4%
FPL P90 (90th-percentile score) 11
Chance of DEFCON points 55.0%
Sorare expected score 59.71
Sorare chance of 70+ 24.7%
Sorare P90 82.08

The pre-GW7 site snapshot also showed him at just 4.6% ownership and £5.0m. He wasn't being promoted because he'd already hauled. He entered the week as a top-ten FPL projection, with a clear defensive-contribution route and a genuine upside distribution.

In the match he produced an assist, a clean sheet and 12 recorded defensive contributions, comfortably over a defender's ten-action threshold. That DEFCON route was central to the prediction: the model had estimated a 55% chance of earning the two defensive-contribution points. His actual 11 FPL points matched his modelled P90. This is a very good example of why we project actions as well as fantasy totals.

And it's not the first time he's featured here. In our GW3 review, Vušković was already on the strong-calls list after a 5.08-point FPL projection became a 12-point return. GW7's 5.50-to-11 follow-up is another encouraging data point — not proof that he'll deliver every week, but a recurring example of FME spotting a less obvious defender whose matchup suits his scoring routes.

Footballers playing a match on the pitch
Generic football action image: Stanley Morales / Pexels. Not a depiction of the players discussed.

Other calls that worked

Erling Haaland was the model's number-one FPL projection: 7.86 expected points, with a 34.2% chance of reaching 10+. He scored 13, thanks to two goals and three bonus points. The premium, obvious pick still has to justify his position at the top of the model; here he did.

In Sorare, Olivier Boscagli went from 61.39 projected to 95.46 actual, having ranked 20th before kick-off. Gabriel Magalhães was another strong high-ranked result: 63.35 projected, 94.60 actual, from pre-match Sorare rank 13. And Jérémy Doku's 74.19 projection came out remarkably close to his 73.90 actual.

These weren't uniformly pinpoint forecasts. They were different examples of a useful pre-match opportunity signal — especially Boscagli and Gabriel, who significantly exceeded their expected scores.

The misses — including our top Sorare player

Bruno Fernandes is the one we cannot dodge. FME ranked him number one for Sorare, projected 79.73, and he returned 34.40. His FPL projection of 5.39 also became only 1 point. A substantial miss on both platforms, and particularly painful at the very top of the Sorare rankings.

Noah Okafor was our second-ranked FPL player at 6.70 expected points and finished with 2. João Pedro carried a 5.47 FPL projection and also scored 2. These are precisely the picks a fair review should put next to the winners.

The model also underestimated several less fancied outcomes. Pedro Neto delivered 15 FPL points from a 3.81 projection, while Kristoffer Ajer exploded to 86.62 Sorare points from 40.39 projected. Those are underestimates, not secretly good calls because there was a small mathematical chance of an unlikely haul.

A probabilistic model will always have surprises, but that explanation is useful only if our probabilities predict the right frequency of surprises over time. That is what the weekly audits need to establish.

Wissa Watch: another quiet return

Those who've followed our earlier reviews will remember Yoane Wissa. In GW3, FME ranked him much lower than the fantasy interest around him, with a 3.24 FPL projection; he scored 1.

The model was similarly unconvinced heading into GW7:

Wissa in GW7 FME projection Actual
FPL points 3.17 1
Sorare score 39.42 38.60

He was 397th on the full FPL expected-points board, with only a 3.2% probability of a 10+ FPL return. The quiet FPL outing was consistent with that cautious assessment. The Sorare projection, meanwhile, was almost exactly in line with his actual score.

This is not a campaign against Wissa, and it isn't a claim that an individual player cannot produce a surprise haul next week. The more useful point is that FME can disagree with the prevailing perception of an asset and make that disagreement testable. We'll keep watching whether those disagreements add value across a much larger sample.

What GW7 tells us — and what it doesn't

The positives: both platforms' average scoring levels were fairly close to projection; the pre-match top-25 survivors outscored the rest; FPL 10+ and 15+ haul totals were broadly in line with their summed probabilities; and a recurring value candidate, Vušković, had another strong week with a particularly instructive DEFCON route.

The negatives: Sorare 60+ and 70+ events happened more often than expected, the individual rank ordering was only moderate, and some of the model's highest-profile picks — particularly Bruno Fernandes and Noah Okafor — disappointed. Pedro Neto and Ajer also show how much individual upside can emerge outside the headline picks.

We are not trying to turn every result into a retroactive success story. Each week adds evidence about what our matchup model sees well and where it still needs calibration.

If you've only recently found us since the website launched, explore the current FPL predicted points, Sorare player projections and DEFCON probabilities. Our model methodology explains how the action-based simulations are built — and why an expected score is a distribution, not a promise.