Multi-step-ahead prediction of IP packet delay variation based on a GARCH model

We propose JitSeer, a long horizon prediction method for the delay variation of IP packets; i.e., jitter. In designing JitSeer, we directed our attention to similarities between delay variation, which is the variation of one-way delay, and volatility in finance, which is the variation of financial asset return. From a time series view, the similarities are that both vary over time, and during some periods they swing quickly while in other periods they stay calm. Based on these similarities, the delay variation is modelled with a GARCH model, which is a time series model originally developed in econometrics to capture the dynamics of volatility. Then the model predicts the future multi-step-ahead delay variation. Experimental results obtained from operational mobile networks show that the prediction accuracy of JitSeer is better than that of conventional methods; its prediction accuracy for the next 10 seconds is 74% over 3G and 88% over LTE.