Comparing forecasting approaches for Internet traffic

Internet traffic is modeled using time series and neural network approaches.FARIMA and ANNs are combined in two different ways for better predictions.A framework for comparison of the different approaches is introduced.Forecasting with a model selected based on non-linearity test is a successful strategy.Alternatively, hybridization between MLP and FARIMA is found to be equally effective. In this paper, we experiment with several different forecasting approaches for Internet traffic and a scheme for their evaluation. First the existence of properties such as Short or Long Range Dependence and non-linearity is explored in order to take advantage of such information and offer a couple of alternatives as forecasting models. The proposed models include FARIMA with Normal and Student's t innovations and two different architectures of Artificial Neural Networks, the Multilayer Perceptron and Radial basis function. Next, we construct a model selection scheme based on the White's Neural Network test for non-linearity or alternatively combine FARIMA and Neural Networks into hybrid forecasting models. The comparison of all suggested approaches is performed using their average position and standard deviation of position when applied to several known datasets of Internet traffic and when the accuracy of forecasts is measured with three different measures. Based on such a data analysis it is shown that hybridization and the selection of a model according to a non-linearity test are more successful as forecasting approaches over all individual models, as well as over other well-known methods such as Holt-Winters, ARIMA/GARCH and FARIMA/GARCH. This result indicates that forecasting approaches which take non-linearity into account lead to better overall forecasts for Internet traffic.