Urban bus services are a vital component of modern cities’ mobility. However, challenges such as traffic congestion, unpredictable weather, and operational incidents often cause delays, negatively impacting service quality and passenger experience. This study introduces GPBus, a genetic programming-based automated machine learning technique designed to predict bus delays. GPBus automates the selection and composition of predictive models while optimizing their hyperparameters. Additionally, the approach integrates transfer learning to reduce computational costs by reusing pretrained models. The proposed model was evaluated using real-world data from the public bus system in Málaga, Spain, spanning several months of operation. Experimental results demonstrate that GPBus significantly outperforms traditional machine learning methods, including Random Forest, Multi-Layer Perceptron, and Support Vector Machines, in terms of prediction accuracy. These results underscore the potential of AutoML approaches to enhance public transport systems by providing accurate and reliable delay predictions, ultimately improving user satisfaction and promoting sustainable urban mobility.