On the Analysis and Predictability of Long-Term Correlated Traffic Patterns in a Wide Area Network

We consider analysis, model identification and predictability of traffic intensity variation patterns for an academic wide area network. We focus on five daily records collected between 2017 and 2021 obtained from the MAWI Working Group Traffic Archive repository. Our results indicate that all considered traffic patterns exhibit long-term correlations characterized by Hurst exponents close to one, superimposed by additional short-term correlations. In turn, they are better approximated by the fractional autoregressive integrated moving average model, that represents both short-term memory effects by the autoregressive and moving average filters, as well as long-term memory effects by an extension of the fractional Brownian motion model. We show explicitly that the above model outperforms the more conventional approach based on the autoregressive model of the same order that implies solely short-term memory, although characterized by various correlation times. Our results including recommendations on the model selection aim at the improved predictability of the traffic fluctuations including emergence of anomalous bursts potentially leading to a better performance of the intelligent network routing algorithms.