By AISEN · Published November 7 2026 · Last updated: November 7 2026
Apps change their features and formulas, so check each app’s own help pages for the latest details.
Quick answer
How sports apps calculate player ratings: human analysts log every on-ball action in a match, software turns those events into stats and probability models such as expected goals, and an algorithm scores each action by its impact. The total is shown as a rating out of 10 and updated live. Sofascore says its rating is calculated by a machine-learning algorithm. FotMob and WhoScored describe algorithms but don’t say whether machine learning is involved, and machine learning also powers supporting metrics like expected goals.
Table of Contents
Key takeaways
- The data starts with people. Opta’s event data is collected in real time by trained analysts, and Stats Perform says computer vision and AI enrich it.
- Sofascore names machine learning. Its help page says its rating is calculated by a machine-learning algorithm. FotMob and WhoScored describe algorithms without saying either way.
- The starting points differ. A 2025 study reports that FotMob and WhoScored start players at 6.0 on a 0 to 10 scale, while Sofascore starts them at 6.5 on a 3 to 10 scale.
- Research shows the systems aren’t interchangeable. The same study, covering 2,190 players, found attacking actions drive ratings most and that WhoScored’s ratings run lower overall.
- Ratings can’t see everything. Off-ball movement, positioning and tactical discipline are mostly invisible to event data.
How sports apps calculate player ratings, step by step
Step 1: People and cameras record the match
FotMob and WhoScored use Opta event data, while Sofascore describes its data as coming from providers it doesn’t name on its rating page. Opta is the data brand of Stats Perform, and, according to Stats Perform’s own glossary, its data is collected in real time by expert analysts, with computer vision and AI used to enrich it. Every pass, shot, tackle and clearance is logged as an event with a time and a position on the pitch, and a typical match produces thousands of them.
This matters for accuracy. A rating can only reflect what was recorded, so anything that isn’t a logged event, such as a run that drags a defender away, is invisible to the formula.
Step 2: Events become stats and probability models
Raw events are turned into stats like pass completion and duels won. Some are turned into models. Expected goals (xG) estimates how likely a shot is to score. Opta Analyst says its xG model uses XGBoost, a machine-learning technique, trained on nearly one million historical shots. FotMob’s help page notes that xG, expected goals on target (xGOT) and expected assists (xA) all come from Opta.
Step 3: Each action gets a weight
This is the heart of a rating: how much each action is worth. Sofascore groups actions into five categories (shooting, passing, dribbling, defending and goalkeeping) and says progressive or riskier passes are rated higher than safe backward ones. It also lists goals, assists, big chances created and last-man tackles as key positives, and red cards, own goals and errors leading to goals as key negatives. WhoScored says every event is weighted by where on the pitch it happened and its outcome.
A good model should value two completed passes differently. A sideways pass in your own half and a pass that splits the defense are both “completed passes,” but the second changes the match far more. Which actions deserve more credit is what makes one app’s rating differ from another’s. The researchers in the study below note that the three systems weight the same statistics differently and that the specific algorithms aren’t public.
Step 4: A starting point sets the scale
Each platform uses its own starting point and range. A 2025 comparative study reports:
| Scale | Starting rating | |
|---|---|---|
| FotMob | 0 to 10 | 6.0 |
| Sofascore | 3 to 10 | 6.5 |
| WhoScored | 0 to 10 | 6.0 |
The study also reports that FotMob requires a player to play at least 10 minutes before receiving a rating. These starting points matter when comparing apps: a 7.5 on one doesn’t necessarily mean the same assessment as a 7.5 on another.
Step 5: The rating updates live
Sofascore says its first rating appears after 10 minutes of play and is then recalculated about 60 times per player during the match. WhoScored says its ratings are calculated live. The final number can still shift slightly after the whistle.
Where AI and machine learning fit (and where they don’t)
The technology behind ratings is more varied than a single formula.
Sofascore says machine learning is involved. Its help page says the rating is calculated by a machine-learning algorithm that assigns a value to each offensive, defensive and goalkeeping action based on its context and impact. The page doesn’t say what the model is trained on or what it is trained to predict, so we can’t tell how it compares with research methods. Sofascore also says it analyzed countless past matches and tweaked its formula.
FotMob and WhoScored describe algorithms. Both draw on large sets of match statistics, but their public explanations don’t reveal enough to tell whether machine learning is used in the rating calculation itself.
Machine learning powers supporting metrics. Opta’s expected-goals model is one example, and Stats Perform says AI also helps collect and enrich its data and, in its Opta Vision product, combine tracking data with event data.
Research shows another approach. VAEP (Valuing Actions by Estimating Probabilities) is an academic method. In the paper that introduced it, the value of an action is how much it changes the team’s chance of scoring minus how much it changes the chance of conceding, with those probabilities estimated by machine-learning models trained on features of the action and its context. A pass that moves the ball into a dangerous area earns credit even if no shot follows. The authors note there’s no objective ground truth for valuing a player. VAEP is research, not evidence that any of these three apps uses it.
The takeaway: football analytics uses a mix of statistical models and machine learning. Don’t assume every rating algorithm is machine learning, and don’t assume every rating works the same way. When an app says “AI-powered,” a good question is what exactly is learned, and from what data.
How FotMob, Sofascore and WhoScored compare
| FotMob | Sofascore | WhoScored | |
|---|---|---|---|
| Data source | Opta (Stats Perform) | “Our data providers,” not named on its rating explainer | Opta |
| What goes in | More than 300 stats per player per match | Hundreds of data points in five categories; described by Sofascore as a machine-learning algorithm | More than 200 raw statistics, weighted by area of the pitch and outcome |
| Scale and start | 0 to 10, starts at 6.0 | 3 to 10, starts at 6.5 | 0 to 10, starts at 6.0 |
| Timing | Needs at least 10 minutes played | First live rating after 10 minutes, then about 60 updates per player | Calculated live |
Sofascore says it has rated players since the 2015/16 season and covers more than 200 competitions, and it claims only about 1 in 3,000 performances earns a perfect 10 (the company’s own figure). FotMob adds percentile rankings that compare a player with others in the same position.
Why the same player gets different ratings
- Different starting points and scales (6.0 vs 6.5, and 0 to 10 vs 3 to 10), so compare ranks, not raw numbers.
- Different weights. The study notes the same statistics are weighted differently by each provider.
- Different inputs. FotMob and WhoScored both use Opta data, but each applies its own model. Sofascore doesn’t name its providers.
- Different timing, while a match is still being played.
What the research found
The best evidence is a 2025 study in the Journal of Sports Sciences, “Comparing player rating systems as a metric for assessing individual performance in soccer.” Per its abstract, the researchers used data from 2,190 players across two seasons in the Premier League, LaLiga and the Bundesliga, and tested 73 performance metrics against ratings from WhoScored, FotMob and Sofascore.
- Attacking actions mattered most: shots on target, key passes and successful take-ons had particularly strong relationships with ratings.
- Defensive actions mattered less, though consistently, for example clearances.
- WhoScored ran lower overall. The model reported a coefficient of -0.20. That’s a statistical model result, not a claim that every WhoScored rating is exactly 0.20 points lower.
- The systems aren’t interchangeable. The authors raise questions about transparency and bias.
The findings concern outfield actions like shots and passes, so don’t assume they apply equally to goalkeepers. The study doesn’t crown a winner. Because there is no true rating to compare against, the fairest tests are transparency, position fairness, stability and whether a rating matches what you saw.
Where ratings mislead
- Off-ball work is invisible. A striker who drags defenders away leaves little in the data.
- Short appearances. A substitute has few events, so one action can swing the rating.
- Team context. A defender on a dominant team may defend less, so his rating can sit in the middle however well he played. This is my reasoning from how event data works.
- Goalkeepers. A quiet match gives few events. FotMob’s “goals prevented” stat, defined as xGOT conceded minus goals conceded, shows how many goals a keeper saved against what was expected.
- A rating isn’t a prediction. It describes what happened, not what will happen.
How to compare the apps yourself in 10 minutes
Treat this as an informal comparison, not a scientific test. Highlights can leave out important defensive and off-ball work, so a full match gives a better basis for judging.
- Pick six players from one match: two defenders, two midfielders, two attackers.
- Write down each app’s rating for them in a table with Player, Position, FotMob and Sofascore columns.
- Check the ranking. Do both apps put the same players in the top two and bottom two?
- Check the spread. Does one app use more of the scale?
- Check the positions. Are defenders consistently lower in one app?
- Check against the match. Watch the full match if you can, and ask whether each rating fits what you saw.
If the apps agree on the ranking but differ on the numbers, that’s the starting-point effect. If they disagree on who played best, you’ve found a real difference in weighting.
Which app should you use?
This is a judgment from the documentation and study, not from daily testing of all three apps.
- FotMob suits you if you want ratings tied to Opta stats and expected-goals numbers, plus percentile comparisons within a position.
- Sofascore suits you if you follow many competitions, since it says it rates more than 200, and like ratings that update during the match from the 10-minute mark.
- WhoScored suits you if you want ratings built from a long list of raw Opta statistics weighted by pitch area and outcome. The study found its ratings run lower overall, which reflects how its scale behaves, not necessarily a more demanding system.
You don’t have to choose. If two apps disagree about a player, look at the underlying stats.
Frequently asked questions
Do sports apps use AI to calculate player ratings?
Sometimes. Sofascore’s help page says its rating is calculated by a machine-learning algorithm. FotMob and WhoScored describe algorithms without saying whether machine learning is involved. Machine learning also powers supporting metrics such as expected goals.
Where does the data behind ratings come from?
FotMob and WhoScored use Opta data. Stats Perform says Opta data is collected in real time by expert analysts and enriched with computer vision and AI. Sofascore says its data comes from providers it doesn’t name on its rating page.
Which app has the most accurate ratings?
None can be proven more accurate, because there’s no true rating to compare against. The research shows the three systems weigh performance differently and aren’t interchangeable.
Why are Sofascore ratings often higher than WhoScored?
One reason is the starting point (6.5 for Sofascore, 6.0 for WhoScored) and the scale (3 to 10 vs 0 to 10). The 2025 study also found WhoScored’s ratings significantly lower overall.
Can I use ratings to predict matches or to bet?
No. Ratings describe what already happened and weren’t designed to forecast results. Never bet more than you can afford to lose.
Does a 10 mean a perfect performance?
Not exactly. It means the model scored the recorded events at the top of its scale. Sofascore says only about 1 in 3,000 performances earns a perfect 10, which is the company’s own figure.
The bottom line
How sports apps calculate player ratings comes down to a pipeline: people and cameras record events, software turns them into stats and models, an algorithm values each action, and the total is shown as a number out of 10. Sofascore says machine learning is part of that, while FotMob and WhoScored don’t say either way. The three systems differ in their starting points, scales and weights, so use one app consistently, compare players within it, and remember that a rating can’t see off-ball work. For more on how apps use data and AI, see our guide to how AI is used in sports apps.
How this guide was researched
I haven’t tested these apps daily myself. This guide is based on each app’s own explanation of its ratings, Stats Perform’s and Opta Analyst’s descriptions of their data and models, and published research, linked in the text. I’ve labeled company claims as such, and noted where the apps don’t publish details. The apps change often, so I’ll update this page and the date at the top when they do.
Also Read: How AI is used in sports
Sources
- Stats Perform: sports data glossary (how Opta data is collected)
- Opta Analyst: what is expected goals (xG)?
- FotMob: help page on ratings and stats
- Sofascore: ratings explained (help center) and rating explainer
- arXiv: Actions Speak Louder Than Goals: Valuing Player Actions in Soccer (VAEP)
- Journal of Sports Sciences (2025): Comparing player rating systems as a metric for assessing individual performance in soccer







