Optimizing energy consumption in buildings, as one of the requirements of sustainable development, requires accurate prediction of heating and cooling loads for the optimal design and operation of energy systems. In this study, with the aim of predicting and comprehensively comparing the performance of different modeling approaches in estimating heat-ing and cooling loads, a dataset of 768 residential buildings with eight geometric and physical properties as input varia-bles and heating and cooling loads as output variables was used. Prior to modeling, the dataset underwent descriptive and exploratory statistical analysis, outlier detection, correlation analysis, and feature selection to identify the most influential predictors and to address multicollinearity among input variables. Six modeling approaches spanning three distinct learn-ing paradigms were then developed and systematically compared: multiple linear regression, decision tree, Random For-est, extreme gradient boosting (XGBoost), support vector regression (SVR), and an artificial neural network whose archi-tecture and hyperparameters were tuned through Bayesian optimization rather than manual trial-and-error. All models were trained and evaluated under an identical train-test protocol, and their performance was assessed using six comple-mentary metrics, including the coefficient of determination, root mean squared error, mean absolute error, mean absolute percentage error, and information criteria. The results showed that the tree-based XGBoost model achieved the highest prediction accuracy among all models studied, with a coefficient of determination of 0.9988 for heating load and 0.9941 for cooling load, followed closely by SVR and the Bayesian-optimized neural network, while linear regression, as the only parametric model, exhibited comparatively lower out-of-sample accuracy. Feature importance analysis, cross-validated across four independent algorithms, consistently identified overall height, relative compactness, and roof area as the dominant predictors of both loads. Additionally, residual analysis revealed a systematic increase in prediction error at peak-load conditions across all six models. These findings demonstrate the strong capability of machine learning meth-ods, particularly ensemble and Bayesian-tuned approaches, to capture the complex relationships between building charac-teristics and energy loads, supporting their use as reliable decision-support tools for energy-efficient building design.