Ensemble Learning with Systematic Hyperparameter Optimization for Urban-Bike-Sharing Demand Prediction

Bike sharing is an established component of urban mobility infrastructure, offering a low-emission alternative to motorized transport for short trips in cities worldwide. Accurate demand forecasting is essential for efficient system operation: it enables better bike redistribution, reduces user wait times, and lowers the operational costs associated with rebalancing. This study evaluated multiple ensemble strategies for hourly bike-sharing demand prediction, comparing bagging methods (Random Forest, Extra Trees), boosting methods (AdaBoost, Gradient Boosting Regressor, Histogram-based Gradient Boosting Regressor), and a Voting ensemble, while systematically investigating the impact of hyperparameter optimization. A repeated hold-out protocol was used, in which the dataset was randomly divided into 80% training and 20% test subsets across 10 random splits; 5-fold cross-validation was applied within each training fold exclusively for hyperparameter tuning, ensuring the test set remained unseen during model selection. Random Search and Bayesian Optimization were compared under identical budgets of 60 configurations per model. Results show that optimization substantially improves all models, with the most pronounced gains for AdaBoost (58% RMSE reduction) and Gradient Boosting Regressor (45% RMSE reduction). A Voting ensemble combining a Random Search-tuned Gradient Boosting Regressor and a Bayesian-optimized Histogram-based Gradient Boosting Regressor achieves the best overall performance (RMSE of 38.48, R2 of 0.955) with the lowest variance among all repeated splits. Feature importance analysis confirms that hour of day and temperature are the dominant demand drivers, consistent with the operational patterns of urban bike-sharing systems. The performance difference between Random Search and Bayesian Optimization is negligible for most models, suggesting that well-designed search spaces allow simpler strategies to achieve competitive results. A controlled comparison conducted under identical experimental conditions shows that the Voting ensemble is statistically equivalent to XGBoost and nominally better than LightGBM, while CatBoost achieves a statistically significant advantage, highlighting it as a strong individual alternative.