The prediction of sports game outcomes has long been a central focus for analysts, bettors, coaches, and sports enthusiasts. Accurate forecasting not only informs betting markets and fantasy sports platforms but also supports decision-making for team managers, sponsors, and media outlets [1] . With the increasing availability of rich statistical data and the continuous advancement of machine learning technologies, sports result prediction has evolved from simple statistical estimations to sophisticated algorithmic systems [2] . Among various sports, the National Basketball Association (NBA) has emerged as an ideal subject for predictive analytics due to its highly structured gameplay, abundant historical records, and global popularity.

A substantial body of research has investigated different methodologies for predicting basketball outcomes. Traditional approaches have included regression-based statistical models and Poisson regression frameworks, which provide interpretable insights but may struggle with complex, high-dimensional data [3] . Subsequent studies incorporated machine learning techniques such as decision trees, support vector machines (SVMs), and neural networks, demonstrating improved predictive accuracy in certain contexts. More recent work has explored deep learning and hybrid approaches, for example integrating neural networks with feature selection or explainable AI methods to enhance both accuracy and interpretability. Comparative evaluations have also been conducted, assessing multiple algorithms under similar conditions to identify performance trade-offs [4] .

Despite these advances, several research gaps remain. Many existing studies focus on a single model or a narrow range of algorithms, limiting the generalizability of findings across different modeling paradigms. The influence of feature engineering and preprocessing techniques, including normalization, encoding, and feature selection, on prediction accuracy has not been systematically compared in the context of NBA game forecasting. Moreover, while some studies evaluate classification-based win/loss predictions, others focus on regression-based score predictions, yet few combine both perspectives in a unified comparative framework [5] .

To address these gaps, this study conducts a comprehensive comparison of multiple machine learning approaches in predicting NBA game outcomes using structured tabular features. The investigated models include Logistic Regression, Support Vector Machines (SVM), Random Forests, Deep Neural Networks (DNN), and the AutoML framework AutoGluon. Both classification (win/loss) and regression (score differential) formulations are examined, allowing for a broader evaluation of model capabilities [6] . Beyond baseline model implementation, we perform systematic hyperparameter tuning for each algorithm to optimize performance, using techniques such as grid search, random search, and automated optimization within AutoGluon. Comparative performance is assessed not only through conventional metrics (accuracy, precision, recall, RMSE) but also via visualization tools, allowing a more intuitive interpretation of model strengths and weaknesses [7] . This study not only bridges the gap between classification and regression approaches under a unified framework but also demonstrates the critical role of preprocessing, systematic hyperparameter tuning, and visual interpretability [8] . By leveraging a large-scale dataset and benchmarking a diverse set of algorithms, our work provides one of the most comprehensive evaluations to date, offering both theoretical contributions and practical guidance for future sports analytics applications.

The dataset comprises 25,796 historical NBA games sourced directly from the official NBA website, containing variables such as game date, team IDs, final scores, shooting percentages, assists, and rebounds, with a binary label for home team victory.

The contributions of this study are threefold: (1) we provide one of the most extensive empirical comparisons of traditional, machine learning, deep learning, and AutoML methods for NBA game prediction; (2) we systematically evaluate the effect of feature engineering and preprocessing on model performance; and (3) we offer practical guidance for selecting predictive models in sports analytics contexts [9] . By identifying the most effective approaches under modern data conditions, this research contributes theoretical and practical knowledge to the growing field of sports analytics and highlights the strengths and limitations of different modeling paradigms when applied to real-world prediction tasks [10] .

This dataset contains detailed historical records of 25,796 NBA games spanning from the 2014 season through the 2022 season, compiled from the official NBA website and supplemented with statistics from Basketball-Reference.com. It covers both regular season and playoff games, providing a comprehensive view of nearly a decade of NBA competition.

Each row represents a single game, identified by a unique GAME_ID, along with the game date (GAME_DATE_EST), season (SEASON), and the IDs of the home and away teams. The dataset includes the final score (PTS_home, PTS_away), shooting statistics (field goal percentage FG_PCT, free throw percentage FT_PCT, three-point percentage FG3_PCT), and core team performance metrics such as assists (AST) and rebounds (REB) for both home and away teams.

A binary label column HOME_TEAM_WINS indicates whether the home team won (1) or lost (0), making the dataset suitable for classification tasks and regression tasks.

This dataset offers a robust foundation for sports analytics research. Its temporal coverage from 2014 to 2022 allows for long-term performance trend analysis, while its granularity enables machine learning experiments in areas such as game outcome prediction, feature importance evaluation, and the effect of feature engineering techniques. The dataset was divided into training and testing part with a test size of 20%. The split was performed randomly, ensuring that approximately 80% of the games were used for model training and the remaining 20% for testing. This approach allows for an unbiased evaluation of model performance.

In this study, a diverse set of machine learning and deep learning models are employed to predict NBA game outcomes, ranging from interpretable linear approaches to complex ensemble and neural architectures.

$\text{accuracy}=\frac{\text{TP}+\text{TN}}{\text{TP}+\text{TN}+\text{FP}+\text{FN}}$

In this study, accuracy is consistently computed using the above formula, where TP and TN denote the number of correctly classified positive and negative samples, and FP and FN represent the misclassified positive and negative samples, respectively.

$R^2=1-\frac{{{\displaystyle \sum }}_{i=1}^n{\left(y_i-{\hat{y}}_i\right)}^2}{{{\displaystyle \sum }}_{i=1}^n{\left(y_i-\bar{y}\right)}^2}$

In linear models, the coefficient of determination $R^2$ is commonly used to evaluate model performance. Here, $y_i$ denotes the true values, ${\hat{y}}_i$ the predicted values, and $\bar{y}$ the mean of the observed values. An $R^2$ value closer to 1 indicates a better fit of the model to the data.

Logistic Regression was selected as a baseline model for the binary classification of home team victories. Logistic regression estimates the probability of a binary outcome using the logistic (sigmoid) function:

$P\left(y=1|X\right)=\frac{1}{1+{\text{e}}^{-\left({\beta }_0+{\beta }^{\text{T}}X\right)}}$

Here, β represents the coefficients, and the output is the probability of the positive class. As a generalized linear model, it estimates outcome probabilities using the logistic function, providing easily interpretable coefficients that indicate the direction and magnitude of feature influence. This simplicity and transparency make Logistic Regression a widely used tool in areas such as credit scoring, medical diagnosis, and sports analytics. In our case, it offered a strong and interpretable starting point, achieving around 77% accuracy with balanced precision and recall.

Linear regression variants, including Lasso and Elastic Net, were applied to the task of predicting continuous score differentials (PTS_home, PTS_away) which were then derive to win/loss outcomes. Positive values indicate a home team win (1), while zero indicate a loss (0). This threshold ensures consistency with the binary labels used in classification models. These regularized linear models were applied to manage multicollinearity and enhance feature selection. Their ability to shrink irrelevant coefficients makes them suitable for identifying the most influential performance metrics, though they may struggle with complex nonlinear patterns in NBA data.

$\hat{\beta }=\arg \underset{\beta }{\min }\left(\underset{i=1}{\overset{n}{{\displaystyle \sum }}}{\left(y_i-{\beta }^{\text{T}}{x}_i\right)}^2+{\lambda }_1{\left|\beta \right|}_1+{\lambda }_2{\left|\beta \right|}_2^2\right)$

This combines L1 and L2 penalties, balancing sparsity and stability when predictors are correlated. While standard linear regression assumes all features contribute equally in a linear fashion, Lasso introduces L1 regularization to encourage sparsity, effectively performing feature selection, while Elastic Net combines L1 and L2 penalties to stabilize predictions when features are highly correlated. These models are commonly used in forecasting, econometrics, and scenarios where interpretability is important. In this study, we progressively improved performance through feature standardization, polynomial feature expansion, and hyperparameter tuning, with a combined Lasso + Elastic Net approach reaching an R² of 0.6346.

Random Forest, an ensemble learning method based on aggregating multiple decision trees, was introduced to capture nonlinear patterns in the data. Random Forest ensemble structure reduces overfitting and highlights feature importance, making it especially valuable in identifying decisive factors like shooting efficiency and rebounds. They are widely used in classification and regression tasks across finance, healthcare, and sports analytics, particularly when feature interactions are complex. In our experiments, Random Forest outperformed single decision trees and provided feature importance measures that highlighted shooting percentage and rebounding as key predictors of game outcomes.

Support Vector Machines (SVM) were also explored for classification.

$\underset{w,b}{\min }\frac{1}{2}{\left|w\right|}^2\text{s}\text{.t}\text{.}{y}_i\left(w^{\text{T}}{x}_i+b\right)\ge 1,\forall i$

SVMs seek an optimal hyperplane that maximizes the margin between classes, and through kernel functions such as the radial basis function (RBF) or polynomial kernels, they can handle nonlinear separations. Known for their effectiveness in high-dimensional spaces and robustness with clear margins of separation, SVMs are used in fields ranging from image classification to bioinformatics. In our case, the best configuration with standardized features achieved an accuracy of 77.49%.

K-Nearest Neighbors (KNN) was employed as a non-parametric, instance-based learning method. KNN offers a simple, instance-based approach that adapts well to local decision boundaries. For NBA game outcomes, it can exploit similarity between games, though its sensitivity to feature scaling and noise makes it less robust than other models.

$\hat{y}=\arg \underset{c}{\max }{\displaystyle \underset{i=1}{\overset{k}{\sum }}1\left(y_i=c\right)}$

It classifies each game based on the majority label among its nearest neighbors in feature space, making no strong assumptions about data distribution. While its simplicity and adaptability to complex decision boundaries are appealing, KNN is sensitive to feature scaling. In our experiments, standardizing features led to a notable improvement, with accuracy increasing from roughly 69% to about 75% [11] .

Deep Neural Networks (DNN), implemented as multilayer perceptrons, were incorporated to capture complex nonlinear relationships in the data.

$h^{\left(l\right)}=\sigma \left(W^{\left(l\right)}{h}^{\left(l-1\right)}+b^{\left(l\right)}\right)$

Here, σ is an activation function and weights are learned via backpropagation. By stacking multiple fully connected layers and training via backpropagation, DNNs can learn hierarchical feature representations. They are a standard choice for structured and unstructured data alike, powering applications in computer vision, natural language processing, and predictive analytics. In this study, optimizer selection and early stopping were used to improve convergence and prevent overfitting, resulting in an accuracy of 77.26% [12] .

We also leveraged AutoGluon, an AutoML framework designed to automate model selection, hyperparameter tuning, and ensembling.

$\hat{y}=\underset{m=1}{\overset{M}{{\displaystyle \sum }}}{w}_m\cdot h_m\left(x\right)$

where $w_m$ are the learned weights of model $h_m$, summing to 1.

AutoGluon evaluates a variety of base learners—such as XGBoost, LightGBM, and neural networks—and combines them into a weighted ensemble that captures complementary strengths. This approach is valuable in scenarios where rapid prototyping and robust performance are required [13] . In our case, the AutoGluon ensemble achieved 77.38% accuracy, surpassing all individual models [14] .

Finally, a Convolutional Neural Network (CNN) was adapted for structured tabular data to explore whether spatial feature relationships could improve prediction accuracy [15] . CNNs, traditionally used in image and spatial data processing, apply convolutional layers to detect local patterns and reduce the number of parameters compared to fully connected networks. They have been widely adopted in computer vision, speech recognition, and bioinformatics. In our experiments, overfitting was mitigated through dropout, batch normalization, and early stopping, resulting in a validation accuracy exceeding 77% and strong generalization performance [16] .

The experiments were conducted using a dataset comprising 25,796 NBA games played between the 2014 and 2022 seasons, covering both regular season and playoff matchups. Game data was sourced from the official NBA website and supplemented with statistics from Basketball-Reference.com.

The prediction task was formulated as a binary classification problem, where the goal was to predict whether the home team won (HOME_TEAM_WINS, where 1 indicates a win and 0 indicates a loss).

Each record in the dataset includes game metadata (date, team ID, season) and home team performance metrics only, including points scored (PTS_home), field goal percentage (FG_PCT_home), free throw percentage (FT_PCT_home), three-point percentage (FG3_PCT_home), assists (AST_home), and rebounds (REB_home).

In the early stage of model experimentation, both home and away team statistics were included as input features. However, as the score differences directly determines the outcome, and the model could easily identify which side had the higher points—a variable that perfectly determines the game outcome. As a result, the model achieved an accuracy of around 98%, which clearly indicated an unrealistic prediction scenario. To avoid this problem and ensure that the model truly learned predictive patterns rather than deterministic labels, we intentionally excluded all away team statistics from the final modeling pipeline. This restriction forces the model to infer a home team’s winning probability solely from its own performance indicators, creating a more authentic and generalizable predictive task.scor

Model Parameter Settings and Loss Functions

The dataset was split into training and test sets using a fixed random seed to ensure reproducibility. Cross-validation was applied for hyperparameter tuning. Model performance was evaluated using accuracy, precision, recall, and F1 score.

Hyperparameters were tuned using different strategies depending on the model type. For tree-based models such as Random Forest, a grid search was conducted with n_estimators ∈ [50, 100, 200], max_depth ∈ [None, 10, 20], min_samples_split ∈ [2, 5], min_samples_leaf ∈ [1, 2], and max_features ∈ ['sqrt', 'log2']. Regularized linear models, including Lasso and Elastic Net, were optimized via cross-validation to select the best α and l1_ratio. For Support Vector Machines (SVM), the kernel type (linear, rbf) and regularization parameter C ∈ [0.1, 1, 10] were tested. In neural network models such as DNN and CNN, learning rate, batch size, and optimizer were tuned manually (e.g., learning_rate ∈ [0.001, 0.01], batch_size ∈ [32, 64]). Visualization tools such as ROC curves, precision-recall curves, and feature importance charts were used to interpret and compare model performance.

While most models achieved competitive results, several factors may explain why certain models produced lower predictive accuracy. K-Nearest Neighbors (KNN) performed the worst (0.7508), as it is highly sensitive to noise and feature scaling, and its reliance on local similarity does not generalize well to large, high-variance datasets. Linear regression variants such as Lasso and Elastic Net also struggled because they assume linear relationships. Though improvements from feature standardization and polynomial expansion, they could not capture the nonlinear interactions that drive basketball outcomes.

Random Forest showed stable but slightly lower performance (0.7685). While it is effective at capturing nonlinearities and providing feature importance, it can overemphasize dominant variables and underweight contextual features. Its ensemble structure also increases variance when the feature space is limited.

Support Vector Machines (SVMs) achieved the highest score (0.7749), benefiting from the high-dimensional representation of standardized features, but they can be computationally expensive and scale poorly with much larger datasets. Deep Neural Networks (DNNs) performed strongly but required careful tuning of optimizers and learning rates to avoid overfitting.

Despite the strong overall performance, this study has several limitations. The models rely only on team-level aggregate data and treat each game as an independent event, which limits their ability to capture player-level dynamics, injuries, and temporal effects such as momentum or fatigue. Contextual factors like coaching strategies or team chemistry are also not represented [18] .

In Figure 11 , it shows that most models perform closely, with Support Vector Machine (SVM) achieving the highest accuracy of 0.7749, followed by AutoGluon (0.7738) and Deep Neural Networks (DNN, 0.7726). These models demonstrate strong generalization and are well suited for complex, high-dimensional data. Traditional models like Logistic Regression (0.77) and Random Forest (0.7685) also remain competitive, while K-Nearest Neighbors (KNN, 0.7508) lags behind due to its sensitivity to data size and noise. Convolutional Neural Networks (CNN, 0.76) showed moderate performance, likely limited by the lack of spatial structure in tabular features. Overall, SVM, AutoGluon, and DNN stand out as the most reliable models for NBA game outcome prediction.

In this study, we conducted a comprehensive evaluation of multiple machine learning models—including regression, classification, ensemble, and deep learning approaches—for predicting NBA game outcomes. The results demonstrate that classification models consistently outperform regression-based models, with Support Vector Machine (SVM), AutoGluon, and Deep Neural Networks (DNN) achieving the highest accuracy, all exceeding 0.77. Feature standardization, polynomial expansion, and regularization significantly improved model performance in linear models like Lasso and Elastic Net, though their explanatory power remained limited. Ensemble methods such as Random Forest and AutoGluon showcased strong generalization abilities, while deep learning models benefited from dropout, batch normalization, and early stopping to prevent overfitting. Overall, models that capture non-linear relationships and leverage ensemble or deep architectures prove more effective for sports outcome prediction using structured tabular data.