Predicting time series of individual trends with resolution adaptive ARIMA

It is important to analyze and predict the time series of traffic trends from the perspective of network operation and management, such as that in fine-grained traffic control, capacity dimensioning, and traffic engineering. However, it is difficult to accurately predict traffic trends because this strongly depends on the time, the type of content, and its popularity. We propose new methods of accurately predicting traffic trends in this paper. Our methods are based on a wavelet transform and the auto regressive integrated moving average (ARIMA) model. We first demonstrate that applying a wavelet transform can improve the accuracy of prediction compared to the original ARIMA model; however, it still has large error due to the fixed time granularity of each resolution. We therefore propose a resolution adaptive ARIMA (RA-ARIMA) model to improve accuracy by changing the time granularity according to the degree of resolution. We demonstrate that by applying it to the real monitored data in a major P2P file-sharing system the normalized mean squared error of RA-ARIMA can be reduced by more than 20% of that of Wavelet-ARIMA. Moreover, we also compared RA-ARIMA with other existing models to prove the accuracy of our proposed method.