Do not sell certainty in Football.
Explore pre-game probabilities for match outcomes, scoring patterns and corners across 140+ competitions. All forecasts are graded after full time.
ML XI app
Inside the app
Explore six core areas, from pre-game probabilities to the public record after full time.
01 / 06
The record comes first
Inspect how ML XI's pre-game probabilities perform over time.
- Every eligible forecast stays in the record, including misses.
- RPS, log loss and ECE measure probability quality.
- Compare results by competition and evaluation period.
- Open a competition to inspect its teams and forecast categories.
Open the Ledger
Rank covered competitions by model accuracy.
Before opening a match
See every team pairing and its saved pre-game call before opening a fixture.
Before opening a settled match
See final scores and graded calls in the list, including correct and missed outcomes.
Inside the Ledger
Probability quality, not just hit rate.
Performance is measured on saved pre-game forecasts, separated by forecast category and sample.
Hover, focus or tap each dotted term.
Ranked Probability Score evaluates the full ordered home/draw/away distribution. Lower is better. Outcome distribution
Log loss strongly penalizes confident probabilities assigned to outcomes that do not occur. Lower is better. Confidence discipline
Expected Calibration Error compares probability bands with how often those events actually occur. Closer to zero is better. Calibration
Benchmark
How accurate is the ML XI model vs Published research?
Most prediction apps quote their own numbers with nothing to compare them to. This page puts ML XI's graded games next to the best published research, market by market, with every source named.
You can verify these numbers yourself, every day, in the app. The free version also shows every graded forecast from the previous three days, HP list included, with each hit and miss marked.
Stated probability vs observed frequency
Each dot is a group of outcomes that carried the same prediction probability before kick-off, home win, draw and away win alike, across all 194,027 games in the ML XI Ledger since 2015. Its height is the observed frequency: how often those outcomes occurred. On the dashed line the two are equal, and the area of a dot is the number of games in the band.
Dot area: games in the band. Whiskers: 95% interval.
Across the 194,027 games, prediction probability and observed frequency are 1.0 points apart on average. Read the dots against the 1:1 line: outcomes given 25% before kick-off occur 26% of the time, those given 34% occur 34%, those given 65% occur 65%. Above 70% the model understates, the one visible bend: outcomes given 74% occur 78%, and those given 83% occur 89%. A probability is not a promise, and an outcome given 70% fails in about 3 games out of 10.
How to read the charts
How often the ML XI model was right, on every game it graded. 194,027 games since 2015 where the Ledger records the market, and 6,956 games since 1 July where it does not. Each bar says which.
The same, counting only the high-probability games the model flags before kick-off. Fewer games, higher hit rate.
The best result anyone has published for the same market, with the paper named under the chart. Hatched means the study is small or picked its games, so it is not a like-for-like figure. Every source is named, with its institution where the work is a thesis.
Six markets, side by side
Match result, double chance and exact score use the full ML XI Ledger: 194,027 games, 145 leagues, 2015 to 2026. Both teams to score, over/under 1.5 and corners are not in the Ledger yet, so they use the published record, 1 July to 17 September 2026. Each bar says which. The "Same league, like for like" strips use the Ledger for that league, 2021 to 2026.
Match result
Who wins, or is it a draw
2015 to 2026
2015 to 2026
2017 challenge test set
Same league, like for like
**Exact score
The precise final score, such as 2-1
2015 to 2026
1 Jul to 17 Sep 2026
Premier League, 2016 to 2018
Same league, like for like
**Double chance
Covering two of the three outcomes, such as home win or draw
2015 to 2026
2015 to 2026
Premier League, 2022
Same league, like for like
**Both teams to score
Do both sides find the net
1 Jul to 17 Sep 2026
4 Jul to 16 Sep 2026
2013 to 2018
Same league, like for like
**Over or under 1.5 goals
Two goals or more, or fewer
19 Aug to 17 Sep 2026
19 Aug to 16 Sep 2026
Premier League and Serie A
Corners
More corner kicks than the line, or fewer
19 Aug to 17 Sep 2026
8 Sep to 16 Sep 2026
Premier League, 2021 to 2023
What this adds up to
- ML XI stands close to or above the best published figures across all 145 competitions, and its pre-game high-probability list scores higher.
- Calibration is the model's strongest result: a stated probability can be taken at face value.
How the grading works
Forecasts are fixed before kick-off and graded after the final whistle. The model is refitted as results arrive and never sees a game's own outcome. ML XI_HP is the subset the machine-learning model sets aside as its highest-probability / high-confidence games.
References
- da Costa, I. B., Marinho, L. B. and Pires, C. E. S. (2022). Forecasting football results and exploiting betting markets: the case of both teams to score. International Journal of Forecasting 38(3), 895 to 909. doi:10.1016/j.ijforecast.2021.06.008. Figures from the first author's 2021 doctoral thesis, Universidade Federal de Campina Grande.
- Wong, A., Li, E., Le, H., Bhangu, G. and Bhatia, S. (2025). A predictive analytics framework for forecasting soccer match outcomes using machine learning models. Decision Analytics Journal 14, 100537. doi:10.1016/j.dajour.2024.100537.
- Reade, J. J., Singleton, C. and Vaughan Williams, L. (2021). Evaluating strange forecasts: the curious case of football match scorelines. Scottish Journal of Political Economy 68(2), 261 to 285. doi:10.1111/sjpe.12264.
- Hubacek, O., Sourek, G. and Zelezny, F. (2019). Learning to predict soccer results from relational data with gradient boosted trees. Machine Learning 108, 29 to 47. doi:10.1007/s10994-018-5704-6.
- Dubitzky, W., Lopes, P., Davis, J. and Berrar, D. (2019). The Open International Soccer Database for machine learning. Machine Learning 108, 9 to 28. doi:10.1007/s10994-018-5726-0. Challenge results table.
- Spann, M. and Skiera, B. (2009). Sports forecasting: a comparison of the forecast accuracy of prediction markets, betting odds and tipsters. Journal of Forecasting 28(1), 55 to 72. doi:10.1002/for.1091.
- Mattera, R. (2021). Forecasting binary outcomes in soccer. Annals of Operations Research doi:10.1007/s10479-021-04224-8. Hit rates on the paper's own selected games, Premier League and Serie A.
- Laurens, M. and Palsson, G. (2023). Forecasting Football Corner Odds. Master's thesis, Lund University. PDF.
- Wilkens, S. (2026). Can simple models predict football and beat the odds? Lessons from the German Bundesliga. Journal of Sports Analytics 12. doi:10.1177/22150218261416681.
- Wang, Z. et al. (2026). WorldCupArena: fine-grained evaluation of language models on football forecasting. arXiv:2607.18084. Preprint.
*Market odds are the prices published by odds compilers, and they do not publish how often those prices are right. Where this page says market odds scored something, a researcher archived the odds, waited for the results, and graded the favourite. The paper that did the grading is named every time.
**Research papers usually study one league over a season or two. ML XI covers more than 140 leagues at once, and some are far less predictable than the Premier League. So where a paper used a single league, we also show ML XI's own result in that same league, under the heading "Same league, like for like".
Pricing
Five days to explore everything.
Then choose monthly or yearly billing. Your app store shows the local price before purchase.
Full access.
- 140+ competitions
- Predictions, live tools and Ledger
- 53 language options
- No ads or trackers
Monthly or yearly Five-day free trial
View availabilityPrice, renewal date and trial terms are shown before purchase and may vary by country or app store.
Help centre
Frequently asked questions
Clear answers about ML XI probabilities, coverage, subscriptions and privacy.
Contact supportPredictions and evaluation
What makes ML XI different from other AI football prediction apps?
ML XI grades its forecasts inside the app and summarizes the results in a clear Ledger. Forecasts are published before kick-off and graded against final results, with hits and misses included in a record of more than 184,000 games across 140+ competitions.
How accurate are ML XI predictions?
Accuracy depends on the forecast type, competition, eligibility rules and evaluation window. ML XI therefore publishes results by probability category instead of one blended headline. The Ledger also reports probability-quality measures, including RPS, log loss and calibration error.
Can AI predict exact football scores?
Not with certainty. AI can estimate the probability of each scoreline, but the quality of those estimates depends on the model. ML XI publishes its score probabilities before kick-off and grades its most likely score after the match.
How does ML XI make its predictions?
ML XI combines machine learning with statistical modelling. Its models are fitted separately to each league using years of results because predictability varies from one competition to another.
Coverage and live matches
Which leagues are covered?
As of September 2026, ML XI already covers 140+ leagues and international competitions across Europe, the Americas, Africa and Asia. Coverage includes top divisions, second tiers, women's leagues, European club competitions, World Cup qualifiers and continental tournaments, with further expansion planned for summer 2027.
Does ML XI cover live matches?
Yes. ML XI provides live scores, attack momentum, match statistics, line-ups and live player ratings. It also includes Goal Hazard, a proprietary ML XI signal that estimates which team is more likely to score next as the match develops.
Subscription and privacy
What does it cost?
ML XI Pro will soon be available on Google Play and the App Store, with monthly and yearly subscriptions after a five-day free trial.
What data does ML XI collect about me?
ML XI requires no account and includes no advertising, analytics or tracking. Preferences stay on your device. Subscription transactions are processed by Google Play or the App Store under their own privacy policies.
Responsible use
Does ML XI guarantee wins?
No, and no honest forecaster can. Football is unpredictable: even a well-calibrated 70% forecast will be wrong about three times in ten. ML XI publishes probabilities and reports their calibration in the Ledger.
Is ML XI betting advice?
No. ML XI publishes statistical estimates and their track record. It does not guarantee outcomes, take wagers or recommend that anyone place a bet.