How accurate is predictapp?
Over 2,189 finished matches in 10 weeks, Trajan, predictapp's soccer model, had the right pick 54.1% of the time, against 51.4% for another data service's prediction model on the same matches. Its confidence levels also closely match real results: when it said about 70%, the pick won 70.2% of the time.
By Daniel, founder of predictapp. Updated 2026-11-01.
How we tested it
Both models were scored on the same matches, using predictions saved before kickoff. Nobody could adjust a number after seeing the result.
- ›2,189 finished matches, with kickoffs from August 28 to November 1, 2026, across 31 competitions.
- ›Scored on the result after 90 minutes.
- ›A model's pick is the higher of its home-win and away-win probabilities.
- ›Trajan's figures are the predictions we published. About 6% of those matches are blended with bookmaker odds, which we explain on the Trajan page.
The results
predictapp came out ahead on all six measures. Pick accuracy is the easy one to read, but Brier score and log loss matter more, because they reward honest probabilities rather than bold guesses.
| Measure | predictapp | Other data service |
|---|---|---|
| Pick correct (higher is better) | 54.1% | 51.4% |
| Pick won or drew (higher is better) | 77.4% | 74.7% |
| Brier score, 3-way (lower is better) | 0.5773 | 0.5917 |
| Log loss (lower is better) | 0.9711 | 0.9915 |
| Calibration error, ECE (lower is better) | 1.84% | 2.14% |
| Largest calibration miss (lower is better) | 6.99% | 7.64% |
- ›Trajan got 58 more matches right, about 5% more.
- ›The two models picked different teams in 381 matches. Trajan was right in 175, the other model in 117, and 89 ended in a draw.
- ›Trajan had more correct picks in 8 of the 10 weeks.
- ›Across the 31 competitions, Trajan was ahead in 20, level in 5 and behind in 6.
- ›On 586 matches, using only Trajan predictions made before the other service's data was collected, a median of 36 hours before kickoff, Trajan scored 56.5% against 53.4%.
- ›For scale: always predicting the sample's average outcome rates gives a Brier score of 0.6434. Trajan is 10.3% better than that, and the other model 8.0%.
Calibration: does 70% really mean 70%?
Calibration asks whether a probability means what it says. If a model says 70% across hundreds of matches, about 70% of those picks should win. It is the most useful test for anyone betting, because you compare that number with the price you are offered.
| Pick confidence | Matches | Stated | Pick won | Won or drew |
|---|---|---|---|---|
| Under 45% | 716 | 40.7% | 43.4% | 69.3% |
| 45-55% | 678 | 49.8% | 51.0% | 75.4% |
| 55-65% | 407 | 59.6% | 57.7% | 82.8% |
| 65-75% | 248 | 69.4% | 70.2% | 87.5% |
| 75% and up | 140 | 81.0% | 84.3% | 95.0% |
How often it expected each outcome
Across all 2,189 matches, Trajan's average probabilities landed very close to what actually happened, including on the draw, the outcome most people forget.
| Outcome | Predicted | Actual |
|---|---|---|
| Home win | 44.6% | 44.7% |
| Draw | 23.5% | 23.3% |
| Away win | 31.9% | 32.0% |
What this test does not prove
A test you can trust has to say where it stops.
- ›It does not show that Trajan beats bookmakers. At prediction time the two were level, about 56% on 530 matches.
- ›Trajan is not fully independent of market data. About 6% of these matches are blended with bookmaker odds.
- ›Trajan is not ahead at every confidence level. At 75% and up, the other model won 87.8% of a smaller set of 49 picks. Trajan won 84.3% of 140.
- ›It says nothing about any single league. The samples are too small.
- ›It covers 10 weeks of soccer. It does not cover the NFL, which has its own record.
Why we do not name the other service
This is a comparison of methods, not an attack on a company. The measures and the method are all here, so you can run the same test on any source you use: save its numbers before kickoff, then score them.
Questions people ask
Is 54.1% a good soccer prediction accuracy?
Yes, for a three-way market where draws happen in about a quarter of matches. Be wary of anyone claiming 80% or 90% over a long run: nobody publishes that with a real, timestamped record.
Does predictapp beat the bookmakers?
We do not claim that. At prediction time the model and the bookmakers were level, about 56% on 530 matches. predictapp shows you a fair price so you can compare it with the one you are offered.
What is the difference between accuracy and calibration?
Accuracy counts how often the pick was right. Calibration checks whether the stated probability was honest. A betting decision needs the second, because the price you get only makes sense against a true probability.
Can I check the results myself?
Yes. The record is public and updates as matches finish, misses included.
See what predictapp says on your match
One free reveal every day, on any match you choose.
Create free accountpredictapp is for adults 18 and over. No model can guarantee a result. Never bet more than you can afford to lose.