Forecasting of traffic load in a live 3G packet switched core network

In this paper we analyze different methods for long term forecasts of packet switched traffic from live 3G networks. The dataset consists of over 400 values, each representing the peak load for a separate day. Four different methods were applied to forecast the increase in traffic, two simple: linear and exponential regression, and two more sophisticated ARMA and DHR. We will show in which cases the sophisticated models deliver a better performance and discuss the question if the gain is significant to justify the increased complexity. We present numerical results for long, e.g., more than 100 day, and short time,e.g., hourly or daily, fitting for our case study based on real traces from a live network. The paper concludes with a benchmark based on the observed mean absolute error.