Parsimonious Fitting of Long-Range Dependent Network Traffic Using ARMA Models

ARMA models are well-suited for capturing auto-correlations of time series. However, in the context of network traffic modeling they are rarely used for their often claimed inappropriateness for fitting Long Range Dependence (LRD) processes. This letter provides evidence that LRD effects can be well approximated by ARMA models; but only the classical fitting algorithms are inappropriate for this task. Accordingly, we propose a novel algorithm, which deploys a multi-scale fitting procedure. It achieves high accuracy up to an arbitrary cut-off lag, yielding parsimonious ARMA models. Our findings encourage a stronger integration of the ARMA framework into the field of network traffic modeling.