Comparing soccer teams that rarely or never meet is a fundamental challenge in sports analytics: clubs play mostly within their national leagues, and national teams within their confederations. We propose a rating framework that places clubs and national teams worldwide on a common probabilistic scale. Building on a recent extension of Glicko with order effects and strength-dependent draw probabilities, we adapt the model to soccer with goal-scoring adjustments, a decomposition of home advantage, extra time, league-aware rating initialization and rules of promotion and relegation that do not inflate the levels of leagues; national teams start from the levels of their confederations. The models are evaluated on about 1.2 million club matches of 252 leagues from 166 countries (2000-2026) and 12.6 thousand national team matches (1998-2026). On held-out seasons, the modified Glicko model is clearly more accurate than Elo and as accurate as score-matrix models built on the goal expectations of a gradient boosting model, including the Dixon-Coles and Markov chain models, while it is substantially more accurate in international club matches. Averaging the predictions of the rating model and the Markov chain model is more accurate than either model alone. For national teams, the model is significantly more accurate than all baselines, especially in the final tournaments, where teams of different confederations meet. The resulting ratings are interpretable and can be used for player evaluation, transfer analysis and match forecasting. The match data, the ratings and the code are publicly available at https://github.com/andreyshelopugin/SportsRatings
Figures & tables
Competition
Matches
Share, %
Clubs, 2000–2026
First divisions
715,021
58.9
Second divisions
408,047
33.6
Domestic cups
55,891
4.6
Domestic super cups
1,530
0.1
Continental and intercontinental competitions
33,606
2.8
Table 1: Number of matches by type of competition
Matches
Train ( < 2021)
Test ( ≥ 2021)
Total
Clubs, domestic
878,816
301,756
1,180,572
Clubs, international
24,222
9,481
33,703
National teams
8,721
3,897
12,618
Total
911,759
315,134
1,226,893
Table 2: Number of matches in the training and test seasons
Group
Features
Venue
venue (+1 home, 0 neutral, −1 away); no spectators; limited number of spectators; home match at a different stadium; match during the COVID-19 pandemic
Stage
extra time possible
Schedule
days since the previous match; matches played in the season; both for the team and for the opponent
Competition
league of the team and of the opponent; tournament; tournament type (first league, second league, cup, super cup, international cup, international super cup, national team qualification, national team final tournament); league level difference (+1, 0, −1)
Table 3: Features of the gradient boosting models (LightGBM and CatBoost) that predict the expected number of goals of a team in a match. Every row of the training set is a pair (match, team), so every match gives two rows. Venue: in the last home matches for a home match and in the last away matches for an away match
Table 4: Fitted parameters of the Markov score matrix, Eq. ( 5 ): the scoring rates of the leading and the trailing team relative to the base rates of the goal model, and the low-score adjustments
Figure 1: Score probabilities (%) for two equally strong teams under independent Poisson goals: (a) both teams expect 1.1 goals, (b) both teams expect 1.6 goals. Home wins are shown in blue, draws (the diagonal) in orange and away wins in purple, darker for more probable scores. The teams are equal in both cases, but the draw probability falls from 29.1% to 23.4% as the expected number of goals grows.
Parameter
Estimate
Description
League
μinit
1330 (253)
initial rating of the league
αcrowd
0.101 (0.068)
home advantage: crowd
αlogistic
0.119 (0.067)
home advantage: travel
α1
0.004 (0.009)
home advantage: strength of the teams
α2
−0.022 (0.019)
home advantage: expected goals
Table 5: Parameters of the model and their estimates on all seasons. League: estimated for every league (253 leagues: 165 first and 88 second divisions); country: for every country (165 countries); global: once for all leagues; fixed: set without estimation by the model. For the league and country parameters, the mean over the leagues and, in brackets, the standard deviation
Figure 2: Home advantage by season: the mean difference between the points of the home and the away team per match (3 for a win, 1 for a draw, after 90 minutes) in domestic league matches, with 95% intervals, and separately for the first divisions of England, Spain, Germany, Italy and France (top 5; about 1,800 matches per season, the interval of about ±0.12 is not shown). The darkness of the band of a season is the share of its matches played during the COVID-19 pandemic (13% in 2019, 93% in 2020, 9% in 2021). The dotted line marks the introduction of VAR in the top European leagues (2017/2018–2019/2020)
Model
Clubs
National
All
Brier
RPS
Skellam (independent Poisson)
0.583899
0.518382
0.583089
0.596645
0.203256
Dixon–Coles
0.583638
0.518363
0.582831
0.596330
0.203203
Markov
0.583569
0.517802
0.582756
0.596266
0.203189
Table 6: Score matrix models on the test seasons (90 minutes): outcome log loss of clubs (311,166 matches), national teams (3,897) and all matches, Brier score and RPS of all matches. All models use the same goal expectations of LightGBM. Bold: the best value in a column
Model
Log loss
Brier
RPS
Δ [95% interval]
All matches (311,152)
Elo
0.591105
0.606114
0.207224
−0.0075 [−0.0079, −0.0070]
Skellam
0.583895
0.597624
0.203627
−0.0002 [−0.0007, 0.0002]
Dixon–Coles
0.583634
0.597304
0.203573
0.0000 [−0.0004, 0.0005]
Markov
0.583565
0.597245
0.203561
0.0001 [−0.0004, 0.0005]
Modified Glicko
0.583646
0.596989
0.203507
–
Table 7: Club matches of the test seasons (90 minutes): outcome log loss, Brier score and RPS. Δ : log loss of the modified Glicko model minus log loss of the model, with a 95% paired bootstrap interval (negative: Glicko is better). Glicko + Markov is the average of the probabilities of the two models. Bold: the best model in a group; in international matches, the modified Glicko model and Glicko + Markov do not differ significantly, and both are bold
Without refit
With refit
Switched off
All
Domestic
Intern.
All
Domestic
Intern.
Home advantage
0.0097
0.0094
0.0189
0.0096
0.0094
0.0181
League-aware initialization
0.0018
0.0005
0.0443
0.0019
0.0006
0.0433
Scoring dependence ( α2 , β2 )
0.0021
0.0021
0.0018
0.0019
0.0019
0.0018
Season growth of ϕ
0.0013
0.0014
−0.0008
0.0013
0.0014
−0.0010
Newcomer offset μnew
0.0003
0.0003
0.0001
0.0003
0.0003
−0.0002
Table 8: Ablation: increase of the test log loss of the modified Glicko model (club matches, 90 minutes) when one component is switched off. Without refit: the other parameters are fixed; with refit: the global and shared parameters are re-estimated without the component
Group
Matches
Actual − Expected
Noise
First season after a transition
Same league
1,061,062
0.001
0.000
Promoted
71,462
0.000
0.002
Relegated
59,648
−0.021
0.002
New in the league
176,116
0.006
0.001
International matches of the test seasons, selected leagues
Table 9: Calibration of the modified Glicko model: actual minus expected score per match (win 1, draw 0.5, loss 0). Transitions: domestic matches since 2010. Leagues: international club matches of the test seasons, predicted with the parameters estimated before 2021: the four leagues with the most matches and the leagues with the largest deviations, and the mean over all 143 leagues with at least 20 matches (the full table is in the repository). The noise level is 0.44/n . I: first division
Rating
Neutral venue
Home of the stronger team
difference
Win
Draw
Loss
Win
Draw
Loss
Regular time
0
36.5
27.1
36.5
44.4
26.6
29.0
10
37.5
27.1
35.4
45.5
26.5
28.0
50
41.8
26.9
31.3
49.8
25.9
24.4
100
47.2
26.3
26.5
55.0
24.8
20.2
Table 10: Outcome probabilities (%) by the rating difference between the stronger and the weaker team, at a neutral venue and at the home stadium of the stronger team (parameters of the English Premier League, a match with the average expected number of goals)
Model
Log loss
Brier
RPS
Δ [95% interval]
All matches (3,637)
Elo
0.510621
0.509617
0.168752
−0.0106 [−0.0145, −0.0063]
Skellam
0.519458
0.519064
0.174164
−0.0194 [−0.0251, −0.0134]
Dixon–Coles
0.519460
0.519140
0.174176
−0.0194 [−0.0252, −0.0134]
Markov
0.518880
0.518611
0.173952
−0.0188 [−0.0248, −0.0127]
Modified Glicko
0.500060
0.496356
0.163617
–
Table 11: National teams, test matches since 2021 (90 minutes): outcome log loss, Brier score and RPS. Δ : log loss of the modified Glicko model minus log loss of the model, with a 95% paired bootstrap interval (negative: Glicko is better). Final tournaments: the World Cup and the continental championships
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 3: The pipeline of the estimation of the club and league ratings
Figure 4: The pipeline of the estimation of the national team ratings
#
League
Rating
SD
n
Int
1500 1750 2000 2250 2500
2016–2025
Best club, rank
1
England I
2441
118
20
214
Arsenal (2)
2
Spain I
2361
133
20
191
Barcelona (3)
3
Germany I
2345
153
18
173
Bayern Munich (1)
4
Italy I
2332
153
20
176
Inter (7)
5
France I
2283
141
18
155
PSG (6)
6
Portugal I
2183
185
18
131
Benfica (17)
Appendix
Table 12: Ratings of the first and second divisions after the 2025/2026 season: the weighted mean of the eight strongest clubs, Eq. ( 22 ). The stripe: the confederation ( UEFA CONMEBOL AFC CAF CONCACAF OFC); SD: the standard deviation of the club ratings (darker: less even); n: the number of clubs; Int: the matches of the clubs of the league against foreign clubs in the last three seasons; the bar: from the median club to the strongest one, the dot: the league rating; the history of the league rating since 2016, relative to the world mean of every season; the strongest club and its world rank. Not shown: 6 leagues without matches against other leagues and without promotion or relegation in the last three seasons (Section 4.5 ) and the Russian leagues
#
Club
Rating
RD
2016–2025
#
Club
Rating
RD
2016–2025
1
Bayern Munich Germany I
2585
54
51
Freiburg Germany I
2215
51
2
Arsenal England I
2560
51
52
Real Sociedad Spain I
2210
52
3
Barcelona Spain I
2538
53
53
Ath Bilbao Spain I
2202
51
4
Manchester City England I
2521
50
54
Celta Vigo Spain I
2200
51
5
Real Madrid Spain I
2490
53
55
Al Nassr Saudi Arabia I
2195
56
6
PSG France I
2482
52
56
Sunderland England I
2194
56
Appendix
Table 13: The 100 strongest clubs after the 2025/2026 season. RD: the rating deviation (darker: less certain); the history since 2016, relative to the world mean of every season
This work reconciles two perspectives on the Elo ranking that coexist in the literature: the practitioner's view as a heuristic feedback rule, and the statistician's view as online maximum likelihood estimation via stochastic gradient ascent. Both perspectives coincide exactly in the binary case (iff the expected score is the logistic function). However, estimation noise forces a principled decoupling between the model used for ranking and the model used for prediction: the effective scale and home-field advantage parameter must be adjusted to account for the noise. We provide both closed-form corrections and a data-driven identification procedure. For multilevel outcomes, an exact relationship exists when outcome scores are uniformly spaced, but approximations are preferred in general: they account for estimation noise and better fit the data. The decoupled approach substantially outperforms the conventional one that reuses the ranking model for prediction, and serves as a diagnostic of convergence status. Applied to six years of FIFA men's ranking, we find that the ranking had not converged for the vast majority of national teams. The paper is written in a semi-tutorial style accessible to practitioners, with all key results accompanied by closed-form expressions and numerical examples.
Predicting a football match before kickoff requires more than knowing past results: a model must use changing information and make a clear prediction before the answer is available. We present WorldCupArena, a dynamic benchmark for language models and deep-research agents. The 2026 FIFA World Cup is its first evaluation, and the same process can be reused for future leagues and cups. Before each match, a model either receives a common evidence package or searches for information itself. It predicts the result and score, likely players and events, match statistics, and the outcome of the competition. After the match, these predictions are compared with the recorded result. We report result accuracy, exact-score accuracy, and a scoreline score that gives some credit when a predicted score is close but not exact, together with scores for the other prediction tasks. Across systems, similar result accuracy can mask larger differences in detailed predictions. Four systems predicted champion Spain, and two of them also recovered the exact final pairing. Compared with betting-market and human-fan baselines, the best system shows only small gains in result and exact-score accuracy, but a clearer gain in Scoreline. New schedules can be added as they begin, allowing the benchmark to evaluate future models without using outcomes that are already known. Code, predictions and evaluation scripts will be publicly released.
Zhaokai Wang, Tianlin Gui, Jiayuan Rao +3
Shanghai Jiao Tong University · Shanghai, China · Nanjing University +4
This paper introduces a score-driven rating system, a generalization of the classical Elo rating system that employs the score, i.e. the gradient of the log-likelihood, as the updating mechanism for player and team ratings. The proposed framework extends beyond simple win/loss game outcomes and accommodates a wide range of game results, such as point differences, win/draw/loss outcomes, or complete rankings. Theoretical properties of the score are derived, showing that it has zero expected value, sums to zero across all players, and decreases with increasing value of a player's rating, thereby ensuring internal consistency and fairness. Furthermore, the score-driven rating system exhibits a reversion property, meaning that ratings tend to follow the underlying unobserved true skills over time. The proposed framework provides a theoretical rationale for existing dynamic models of sports performance and offers a systematic approach for constructing new ones.
Vladimír Holý, Michal Černý
Prague University of Economics and Business Winston Churchill Square 1938/4, 130 67 Prague 3, Czechia