Can We Achieve Better Wireless Traffic Prediction Accuracy

Wireless traffic prediction is of great significance in wireless networks and communication systems, and has been widely studied. However, the theoretical limits of this prediction remain open, especially on whether the theoretical limit is valid and whether it can be approached by the actual prediction algorithm. This article performs voice traffic predictability analysis based on real entropy and five mainstream prediction models to answer this unsolved question. We have two findings in this research: Predictability can be considered as the theoretical guidance on prediction accuracy, and the prediction performance of state-of-the-art models and algorithms still have gaps to the prediction bound. Using two kinds of real-world voice traffic data, the prediction performance of the most frequent visit models - multi-order Markov chains, multi-order diffusion kernel models, multi-order SVM, and LSTM - are obtained and discussed. The most frequent visit model and LSTM model have the best prediction effects on users' traffic and base stations' Erlang, respectively. This work enriches the understanding and application of predictability theory and sheds new insight on voice traffic prediction issues.