Projected Scores on Sorare Inside:The New Model in Numbers

From today, the Projected Scores on Sorare Inside come from a new SoProb model. On the most recent Gameweek it was clearly more accurate than the previous projection, its values matched the real scores across the whole range, and in the historical tests its top picks reached 70+ far more often.

From today, the Projected Score (PS) on Sorare Inside comes from a new SoProb model, built from scratch. It replaces the previous projection completely. This article shows with numbers what changes: on the most recent Gameweek, and over 76 Gameweeks of historical tests.

At a glance

The most recent Gameweek (GW 720, 4,006 starters with 45+ minutes):

76 historical Gameweeks:

What the Projected Score means

The Projected Score is what the player scores on average in this match, as if the match could be replayed many times. Like before, it assumes that the player starts and plays at least 45 minutes.

It is calibrated: players projected at 60 actually score about 60 on average. Because it is an average, most values sit between 40 and 60, and extreme outcomes are smoothed out. A projection of 65 is already a very strong value; in the tests only 0.25% of all projections were above 70.

The most recent Gameweek: GW 720

For the 4,006 starters who played at least 45 minutes in the matches of GW 720 played so far, I compared both projections, exactly as they stood before kickoff, with the real Sorare scores.

The new projection missed the real score by 13.8 points on average, the previous one by 15.0. The previous projection was even less accurate than the player's L10 average (14.4) and than giving every player the same value (14.7).

Do the numbers mean what they say?

A projection is only useful if a 60 really means about 60. The same 4,006 starters, grouped by their projection, with the average projection of each group and the average real score:

The previous projection was too extreme in both directions. Players it projected low scored far more than expected, and players it projected high far less: a previous 64 meant 55 in reality, a previous 77 meant 58. For every point the previous projection went up, the real score went up by only 0.35 points.

The new projection stays within about one point of the real average in every group from below 40 up to 70, and for every point it goes up, the real score went up by 0.88 points. Above 70 there were only 20 new projections in GW 720, too few for a firm statement; in the historical tests, starters projected between 70 and 80 scored 73.0 on average.

Picking the best players

Projections are mostly used to pick players, so the ranking matters most. In GW 720, per position, the starters with the highest projection:

One Gameweek can be lucky or unlucky, so the model was also tested over a long period. I compared it with the projections previously shown on Sorare Inside, on the same matches across 76 historical Gameweeks, looking at the top five players per Gameweek, competition type and card position:

The top five by the new projection reached 70+ 12.5 percentage points more often (95% interval: 10.4 to 14.6) and 80+ 9.6 points more often (7.9 to 11.4). Starters it projected between 70 and 80 scored 73.0 on average; previous projections in the same band came in at 56.6.

Why the previous projection fell short

The previous projection fitted its own set of weights for every player, from that player's games alone. Half of the players had 15 or fewer matching games, so these weights picked up a lot of chance, and the projections swung further than the real scores ever did. That is what the grouped table shows: too low at the bottom, too high at the top.

The new model learns from all players at once how to read a player's history and the matchup. Every single projection still uses only that player's own data:

• The player's own history at the card position, overall and split by competition type and home or away, and at the current club, so that after a transfer it counts how the player performs there
• The player's role: how regularly they have started recently
• The odds for exactly this match: win probability, expected goals for and against, clean-sheet probability, the season possession averages of both teams and, where available, the player's own scorer and assist odds

There are no shortcuts based on league or position: a defender is not projected higher because defenders in that league usually score well.

The model is always tested on Gameweeks it has never seen: it learns only from earlier Gameweeks and then projects the following ones, step by step through the history. SoProb recalculates the projections twice a day with the latest odds and sends them to Sorare Inside; at kickoff they freeze.

Limitations

• Football stays football. A strong projection does not protect against an early substitution, a red card or a missed chance.
• The Projected Score assumes that the player starts and plays at least 45 minutes. Always check the starting probability, the expected minutes and injury information as well.
• The odds are only as good as the market at the time of the calculation; matches without odds yet are projected from the player's own games and are less reliable.
• GW 720 is a single Gameweek. The historical tests over 76 Gameweeks are the more robust comparison, and both point the same way.

Conclusion

The new model is more accurate, its values mean what they say, and it picks clearly better players than the previous projection: in GW 720 and over 76 Gameweeks of tests. From today it powers every Projected Score on Sorare Inside.