Ali Daeihagh, Dhruv Belai, Noah Smith
Khoury College of Computer Sciences
Northeastern University
Boston, MA
daeihagh.a, belai.d, smith.noah @northeastern.edu
October 29, 2024
In recent years, sports betting has rapidly grown into one of America’s largest industries, allowing individuals to place bets on a variety of sports. The introduction of predictive models has proven highly beneficial in enhancing decision-making for betting on game outcomes, as these models provide valuable insights into game dynamics, enabling bettors to make more informed wagers. Beyond betting, predictive models are also highly useful for team front offices, which can use them to identify key factors that contribute to their team’s performance. By analyzing these insights, teams can pinpoint areas for improvement in practice, focusing on the elements that most significantly impact game outcomes.
Predictive modeling also influences sports journalism, shaping discussions, debates, and narratives. Talk shows and media increasingly rely on these predictions to create engaging content and capture audience interest. With these developments, there is a clear demand for accurate models to predict sports outcomes. While past models have shown success, there remains significant room for improvement. Efforts to enhance predictive accuracy may involve refining the set of key features considered and exploring various model architectures to achieve better precision.
Applying machine learning to determine the outcomes of sports games is not necessarily a new endeavor. Decades of research have introduced diverse methodologies to forecast various aspects of sports. For example, within the NBA alone, machine learning has been used to predict playoff contenders, All-Star selections, game outcomes, as well as to determine the viability of the “Hot Streak Fallacy” [7]. While the specific approach varies based on the prediction goal, it is crucial to recognize the range of methodologies available to address different aspects of sports prediction.
Traditionally, sports prediction has been approached as a classification problem, though numeric methods have also been employed [2, 1]. For instance, Delen et al. (2012) used regression-based models to predict NCAA bowl game outcomes, employing neural networks, decision trees, and support vector machines. However, when comparing these models with classification-based approaches, they found that classification models provided more accurate game outcome predictions. This preference for classification is widely supported across research groups, and there is a general consensus that sports predictions are best suited to classification approaches [1].
In classification-based sports prediction, a variety of models have been explored. These range from K-Nearest Neighbor clustering, support vector machines, and naive Bayes to neural networks and decision trees [7, 1, 5]. Each model has its own strengths and weaknesses, yet neural networks are among the most commonly used and accurate methods in sports prediction [1]. Neural networks have proven effective at creating precise classification models across other fields that rely on extensive input data [3].
NBA game prediction models have employed a wide range of input features to improve predictive accuracy. Team statistics have been traditionally used, with Thabta et al. (2019) achieving 83% accuracy in game predictions based on 20 team-based statistics per matchup. Among the possible team features to include when predicting, defensive rebounds, three point percentages, free throws made, and total rebounds are the most influential for improving model accuracy [6]. It is worth noting that player based statistics have not always been utilized when developing these models. This could be detrimental as it overlooks the impact that individual players may have on the game. There are many instances in NBA games where an individual player can single handedly take over and dominate a game, which makes the inclusion of player based statistics all the more important. In addition, many studies miss out on the temporal aspect of team and player statistics when developing and predicting their models. This could be particularly important because teams often hit a period of momentum or a winning streak within their season in which they may be playing better than they were before. Considering this could offer marked improvements in predictive accuracy [5].
Khanmohammadi et al. (2024) have built on these areas for improvement by introducing a recurrent neural network that integrates both player and team statistics to predict NBA game outcomes. This network's recurrent structure allows it to capture the temporal dynamics of team and player performance, while the addition of player-based statistics addresses previous models' shortcomings. Their model achieved an 82% accuracy in predicting NBA game outcomes [5]. Drawing on prior research, we aim to further investigate the interplay between player and team statistics and explore how combining both within a neural network might enhance prediction accuracy in comparison to the results achieved by Khanmohammadi et al.
Our project centers around the extensive NBA Database [4] available on Kaggle, a resource that spans over 65,000 games, capturing every NBA matchup since the league’s very first season in 1946-47. This dataset not only covers comprehensive information on games, teams, players, and referees but also includes an extensive play-by-play record with over 13 million rows, providing a detailed breakdown of in-game events. This wealth of data offers a rich foundation for analyzing trends, extracting insights, and making data-driven predictions for current games.
With over 15 interconnected tables, each offering unique perspectives on NBA games, we will selectively utilize key features across these tables to build a predictive model with greater accuracy and depth than previous attempts. By focusing on nuanced, relevant aspects of gameplay and player performance, this approach aims to push the boundaries of game outcome forecasting, leveraging both historical patterns and real-time analysis for an edge in predictive power.
Our project aims to predict the outcomes of NBA games for the 2024-2025 season—specifically, predicting game results in terms of win or loss. With access to a robust dataset covering historical team, player, and play-by-play data, we will leverage this resource to build a model that accurately reflects the dynamics of live NBA games. Our approach combines aggregated team and player statistics with sequential play-by-play information, enabling us to capture both static trends and real-time in-game shifts.
For team statistics, we will compile metrics like points per game, offensive and defensive ratings, assists, rebounds, and turnovers for each team. Given the scale of our data, our preprocessing strategy includes aggregating these metrics across multiple seasons, with normalization to account for evolving play styles and rule changes. These metrics will give a high-level view of team performance and tendencies, providing reliable features that will complement real-time game data.
For player statistics, we’ll extract individual contributions such as shooting percentage, assists, blocks, and fouls. These will then be aggregated to create team-level impact features that reflect the influence of each player’s contributions on game outcomes. To capture player availability realistically, we will weight player contributions based on average playing time and adjust for any injuries or absences, thus preserving the consistency and accuracy of our team-level features. This aggregation will allow us to quantify each team’s current roster strength, factoring in active versus injured players for a more accurate prediction of game-day performance.
The play-by-play data is structured sequentially, capturing each action in chronological order. We will utilize this format to build a comprehensive view of game flow, extracting features such as possessions, scoring runs, turnovers, and fouls, which are key to understanding momentum shifts. For efficient modeling, we will employ feature engineering to standardize this data across games by padding or truncating sequences to a fixed length, allowing the model to process and interpret each game sequence effectively. This standardized approach ensures that the RNN processes sequences of comparable structure, making training more stable and generalizable across games.
To uncover underlying patterns, we plan to perform exploratory data analysis (EDA) on our aggregated features. Key visualizations will include correlation matrices, which will help us pinpoint relationships between features and game outcomes, as well as trend analyses of play-by-play sequences to identify impactful in-game events. This analysis will guide our feature selection, helping us identify and prioritize features with strong predictive value. Additionally, EDA will provide a solid foundation for hypothesis testing and further validation of our model design.
Our model architecture will integrate both static and dynamic data using a hybrid approach:
- RNN for Sequential Play-by-Play Data:
- To capture the temporal dependencies in play-by-play sequences, we will implement a Long Short-Term Memory (LSTM) network. LSTMs are particularly suited for handling the sequence of events in a game, as they can learn dependencies between plays that span over different timescales.
- Each action within the sequence will be encoded into a vector representing event types (e.g., scoring play, turnover, foul), and the LSTM will output a vector summarizing the impact of these events on the overall game momentum and likely outcome.
- Tree-Based Model for Team and Player Statistics:
- For processing team and player statistics, we will employ a tree-based model such as XGBoost or Random Forests. These models handle structured, tabular data effectively and are well-suited to learning from aggregated statistics, capturing non-linear relationships and interactions among features.
- This tree-based model will allow us to incorporate static data reliably, providing a prediction layer that represents pre-game factors such as team strength and player availability.
- Combining RNN and Static Data Outputs:
- After training both the LSTM and the tree-based model, we will combine their outputs by concatenating the feature vectors from each model. This hybrid vector will then pass through a dense layer that learns how to integrate the insights from both models, capturing the combined influence of pre-game statistics and in-game events on the final outcome.
- The dense layer will enable the model to weigh the relative importance of static and dynamic data, providing a holistic prediction that considers both game context and team dynamics.
Predicting game outcomes on a week-by-week basis could be interesting, but separating training and test data by a historical cut-off date might yield a more consistent evaluation. By using this approach, we can start evaluating our model’s accuracy from a fixed point in time and measure its performance on a more substantial dataset.
To enhance our evaluation, we can incorporate the final game scores to adjust the results. Including the score in our prediction accuracy calculation would allow us to account for discrepancies between predicted and actual outcomes more effectively. For instance, if the model predicts a team will win but they lose, the evaluation should penalize the model more heavily if the loss was by a large margin than if it was by a close score.
Additionally, we’ll compare results from different models and parameter combinations to identify the features that play a significant role in predicting game outcomes. In the NBA, performance can be influenced by recent results due to factors like fatigue, team morale, and momentum from winning or losing streaks. We will assess our model’s ability to capture these temporal patterns by applying techniques such as sliding window averages. We’ll also analyze specific game samples from unique periods like pre-season, tournament play, and playoff games to gauge the model’s adaptability across different contexts.
We expect our model to predict game results with a high degree of accuracy, although we recognize that some outliers may arise due to factors not captured in our dataset. There is a level of stochasticity which is one of the things that makes the NBA so captivating to watch. Nonetheless, we anticipate that the model will accurately favor higher-ranked teams in many cases, reflecting the tendency for such teams to win.
Our goal is to identify a set of key features that consistently contribute to accurate predictions. We expect that more complex models, which incorporate both static and dynamic information, will outperform simpler models with a limited feature set.
While we aim to include a wide range of features, we may ultimately select only a few high-impact features. Ideally, we would include data on individual player performance, but due to the data’s complexity and potential inconsistencies, we may need to rely on standard features that are more reliably available.
After some discussion, we have devised the following list of tasks to be completed: data aggregation and preprocessing, exploratory data analysis, model development, hyperparameter tuning and cross-validation, testing and evaluation, and documenting results and interpreting findings. We have decided to divide the tasks among team members, with group collaboration as needed. Dhruv will work on data aggregation and preprocessing, while Noah will take on exploratory data analysis, and Ali will focus on model development and hyperparameter tuning.
From there, we all plan to work together on testing, evaluation, and contributing to the results and interpretation of findings. Each individual's task is an essential part of the planned pipeline for this project, making each group member’s contribution both important and necessary. These tasks are tentative, and we anticipate some level of shared contribution within each task.
[1] Bunker, Rory P., and Fadi Thabtah. “A Machine Learning Framework for Sport Result
Prediction.” Applied Computing and Informatics, vol. 15, no. 1, Jan. 2019, pp. 27–33,
www.sciencedirect.com/science/article/pii/S2210832717301485,
https://doi.org/10.1016/j.aci.2017.09.005.
[2] Delen, Dursun, et al. “A Comparative Analysis of Data Mining Methods in Predicting
NCAA Bowl Outcomes.” International Journal of Forecasting, vol. 28, no. 2, Apr. 2012,
pp. 543–552, https://doi.org/10.1016/j.ijforecast.2011.05.002.
[3] Mohammad, Rami M., et al. “Predicting Phishing Websites Based on Self-Structuring
Neural Network.” Neural Computing and Applications, vol. 25, no. 2, 21 Nov. 2013, pp.
443–458, https://doi.org/10.1007/s00521-013-1490-z.
[4] “NBA Database.” Www.kaggle.com, www.kaggle.com/datasets/wyattowalsh/basketball.
[5] Reza Khanmohammadi, et al. “MambaNet: A Hybrid Neural Network for Predicting the
NBA Playoffs.” SN Computer Science, vol. 5, no. 5, 10 June 2024,
https://doi.org/10.1007/s42979-024-02977-0. Accessed 17 Sept. 2024.
[6] Thabtah, Fadi, et al. “NBA Game Result Prediction Using Feature Analysis and Machine
Learning.” Annals of Data Science, vol. 6, no. 1, 3 Jan. 2019, pp. 103–116,
https://doi.org/10.1007/s40745-018-00189-x.
[7] Wang, Jingru, and Qishi Fan. “Application of Machine Learning on NBA Data Sets.”
Journal of Physics: Conference Series, vol. 1802, no. 3, 1 Mar. 2021, p. 032036,
https://doi.org/10.1088/1742-6596/1802/3/032036.