Leakage-aware pre-match prediction of professional volleyball outcomes: temporal validation of machine-learning models against an Elo baseline
This study demonstrates that while machine learning models can moderately predict professional volleyball outcomes when rigorously tested against data leakage and temporal constraints, a simple Elo rating baseline remains highly competitive, suggesting that complex algorithms offer limited additional value for pre-match decision support.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Predicting the outcome of a professional sports match is a task that blends intuition with hard data. In the world of competitive volleyball, teams are constantly analyzed for their strengths in serving, attacking, and blocking. For decades, researchers have looked at statistics gathered after a game ends to understand why one team won and another lost. However, there is a crucial difference between explaining a result after the fact and predicting it before the first ball is served. A statistic like the number of points scored in a match is a perfect description of what happened, but it cannot be used to forecast what will happen, because it does not exist until the game is over. True forecasting requires using only the information available right before the match begins, such as a team's past performance, their recent form, and how many days they have had to rest. This distinction is vital for anyone trying to build a reliable model to guess the future, as using data that shouldn't be available yet creates a false sense of accuracy.
A recent study by Xiaopeng Li from Chengdu Sport University tackled this challenge by examining professional volleyball matches in Poland's top league, known as the PlusLiga. The goal was to see how well machine learning models could predict match winners when strictly forbidden from using any information that would not be known before the game started. The researcher worked with a public dataset containing thousands of match records, but the first step was not to build a computer program, but to clean the data. Upon closer inspection, the records from the earlier years of the dataset contained impossible numbers, such as a team recording more "aces" (perfect serves that the opponent cannot return) than total serves attempted. Because these errors could not be fixed without guessing the original rules of the data collection, the researcher made the difficult decision to discard over two thousand of these older records. The final analysis relied on 628 matches from the 2020/21 through 2022/23 seasons, a period where the data was clean and reliable.
To test the predictive power of different methods, the researcher split this clean data into two groups. The first group, covering the 2020/21 and 2021/22 seasons, was used to train various computer models. The second group, consisting of matches from the 2022/23 season, was kept completely hidden until the very end. This approach mimics real life: a model learns from the past and is then tested on the future to see if it actually works. The study compared five different types of machine learning algorithms, ranging from simple statistical methods to complex neural networks, against two basic benchmarks. One benchmark was simply guessing based on how often teams won in the past, while the other used the Elo rating system, a transparent method that updates a team's strength score after every single match based on who they beat and how.
The results showed that predicting professional volleyball is possible, but it is far from perfect. The most successful model was a straightforward statistical method called Logistic Regression, which correctly predicted the winner about 72 percent of the time on the unseen future data. This was a solid result, but it was not a breakthrough. In fact, the simple Elo rating system, which only looked at the difference in team strength without any complex calculations, performed almost as well, getting the winner right about 70 percent of the time. The more complicated machine learning models, including those designed to find intricate patterns in data, did not offer a significant advantage over the simpler methods. In some cases, the complex models were even less stable, performing well during training but failing to maintain that success when faced with new matches.
A deeper look at what the models were actually using to make their decisions revealed a surprising truth. The most important factor by far was the difference in the teams' Elo ratings. While the models had access to fourteen different statistics about a team's recent form, such as their serve accuracy or how many points they scored in their last five games, these details added very little extra value. The computer models essentially relied on the simple fact that a stronger team is more likely to win, and the extra technical data did not change that prediction much. This suggests that the information contained in a team's recent match history is already largely captured by their overall strength rating. The study also found that the models were good at ranking the likelihood of a win, but they were not perfect; there was still a significant amount of uncertainty in every prediction.
The study concludes that while it is possible to forecast professional volleyball outcomes with moderate success, the path to doing so requires strict discipline about what information is allowed. The research explicitly argues against the idea that more complex computer algorithms are automatically better for this task. Instead, it shows that a clear, transparent measure of team strength is often the most powerful tool available. The findings support the use of these models as a cautious aid for pre-match analysis, perhaps helping scouts or fans understand the likely difficulty of a matchup. However, the study warns that these models should not be used for live, in-game decisions or to give specific coaching advice, as they cannot account for the dynamic changes that happen during a match. Furthermore, the results are specific to this one league and cannot be assumed to work for other sports or different levels of competition without further testing. The work serves as a reminder that in sports forecasting, knowing what you do not know is just as important as the prediction itself.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.