Multistep ahead prediction of long-range dependent teletraffic

While long-range dependent (LRD) aggregate network traffic can always be approximated by an autoregressive processes AR(p), the order p required to achieve a good approximation may be so large as to make parameter estimation extremely difficult. Besides, it is necessary to overcome the fundamental problem of the order selection method in this case, as the order is usually unknown. An alternative approach is to postulate that LRD teletraffic follows an ARFIMA (AutoRegressive Fractionally Integrated Moving Average) process, with which we can obtain multistep ahead forecasts via a finite dimensional state-space (truncated) representation of the estimated ARFIMA model. Our results, which were obtained using a genuine teletraffic trace and a simulated series, indicate that the ARFIMA approach can provide a similar or better forecasting performance than high-order AR models.