A comparison of information criteria for traffic model selection

Abstract

Traffic modelling is a core component of network planning and engineering. Although good models are approximations of reality, they are very useful in various network applications. However, traffic modelling is often done in an ad hoc manner, guided only by the experience of the model designer. In this paper, we propose the use of information criteria, such as the Akaike Information Criterion (AIC), to systematically choose models. We study these criteria on Frequency, Frequency + Spike, and Wavelet models of the network traffic to select the best of these. However, there are many alternative information criteria, which give different results. We found that the Bayesian Information Criterion (BIC), and Minimum Description Length (MDL) provided better models than the (perhaps) more commonly used AIC and corrected AIC for network traffic modelling. Interestingly, we found that fancier models, such as Wavelet models, may reduce prediction accuracy, so simple frequency-based models are preferable.