Deep learning-based traffic prediction models require vast amounts of data to learn embedded spatial and temporal dependencies. The inherent privacy and commercial sensitivity of such data has encouraged a shift towards decentralised data-driven methods, such as Federated Learning (FL). Traditional machine learning models capture spatial and temporal relationships in centralised data. In reality, traffic data is likely distributed across separate data silos owned by multiple stakeholders. In this work, a cross-silo FL setting is motivated to facilitate stakeholder collaboration for optimal traffic flow prediction applications. This work introduces FedTPS, a novel framework which augments each client’s local dataset with synthetic data generated by a federated diffusion-based model. The proposed framework is evaluated on three large-scale real-world datasets and assessed against various generative and spatio-temporal prediction models, including a novel prediction model we introduce which leverages Temporal and Graph Attention mechanisms to learn embedded Spatio-Temporal dependencies. Experimental results show that FedTPS outperforms several FL baselines with respect to global model performance.