<p>Accurate forecasting of bike-sharing demand is essential for optimizing fleet management and improving user satisfaction. Traditional models often struggle to capture the complex spatiotemporal dependencies required for high-quality predictions. This study proposes the ApexBoost Regression (ABR) model, which integrates decision tree regression with gradient-based boosting to enhance the modeling of intricate temporal and spatial patterns in bike rental behavior. The model was applied to the Metro Bike Sharing dataset, where the data were partitioned chronologically: the full year of 2023 (36,312 data points, approximately 80%) was used for training, and the fourth quarter of 2024 (9,118 data points, approximately 20%) was reserved for testing. To improve predictive performance, we conducted extensive feature extraction, including spatial variables (longitude, latitude, and location cluster), temporal indicators (hour, day of week, day of month, is_morning, is_evening), and cyclical encodings (sine and cosine transformations of hour). We then employed a feature selection algorithm (FSA), guided by both correlation analysis and feature importance metrics, to retain the most informative and non-redundant variables. Our exploratory data analysis (EDA) revealed important usage patterns such as peak rental times across seasons and day types, supporting feature selection choices.Comparative evaluation against state-of-the-art models-including XGBoost and Random Forest-demonstrates the superior performance of ABR. Prior to hyperparameter tuning, the model achieved a mean squared error (MSE) of 3.0501. After optimization using particle swarm optimization (PSO), the MSE was further reduced to 3.0164, confirming the effectiveness of both the proposed model and the optimization strategy.</p>