Mobility prediction is becoming one of the key elements of location-based services. In the near future, it will also facilitate tasks such as resource management, logistics administration and urban planning. To predict human mobility, many techniques have been proposed. However, existing techniques are usually driven by large volumes of data to train user mobility models computed over a long duration and stored in a centralized server. This results in inherently long waiting times before the prediction model kicks in. Over this large training data, small time bounded user movements are shadowed, due to their marginality, thus impacting the granularity of predictions. Transferring highly sensitive location data to third party entities also exposes the user to several privacy risks. To address these issues, we propose MobiDict, a realtime mobility prediction system that is constantly adapting to the user mobility behaviour, by taking into account the movement periodicity and the evolution of frequently visited places. Compared to the existing training approaches, our system utilises less data to generate the evolving mobility models, which in turn lowers the computational complexity and enables implementation on handheld devices, thus preserving privacy. We test our system using mobility traces collected around Lake Geneva region from 168 users and demonstrate the performance of our approach by evaluating MobiDict with six different prediction techniques. We find a satisfactory prediction accuracy as compared to the baseline results obtained with 70% of the user dataset for majority of the users.