Online Transfer Learning (TL) allows knowledge to be learnt from a data rich source domain to aid predictions in an online target domain. However, when all domains are online, and a data rich source domain does not exist, we must determine what to transfer, how to combine transferred knowledge, and whether to transfer knowledge. To ensure the feasibility of online TL methods in real-world applications, they should not only aid predictions in receiving domains, but should consider the communication and computational overheads of knowledge transfer. To address these challenges, this thesis presents methods for online TL when all domains are online, which are evaluated using synthetic and real-world regression-based datasets. First, the BOTL framework is introduced, which enables knowledge transfer to be conducted bi-directionally between online data streams, where knowledge is transferred in the form of predictive models, and combined using an OLS metalearner. Second, two methods of selecting a relevant yet diverse subset of transferred and locally learnt models are presented, namely parameterised thresholding and conceptual clustering. These approaches help to prevent over_tting when the number of models transferred is large in comparison to the window of available data. To reduce the computational overhead of selecting subsets of models, a static diversity metric is introduced, which estimates the conceptual similarity between models using the Principal Angles (PAs) between their underlying subspaces. Third, two methods for determining whether to transfer knowledge are presented, namely IdDT and IdCS, which maintain comparable predictive performances to when all models are transferred, while reducing the number of models received in each domain by 47:1% and 30% respectively across the experiments conducted for this thesis.