Modeling and Forecasting Web Traffic

Understanding traffic flows is always important, for observation of performance, and for planning. Various types of telecommunications traffic have been shown to be chaotic, and hence awkward to simply describe even if averaged over a period of time. Being able to forecast accurately would therefore be quite valuable - to allow for sudden demands on the system, or to take preemptive action through aggresive prefetching. The plot below shows some hourly Web hit data.

The superimpostion of scales and of fluctuations or cycles is potentially very valuable. It allows the underlying signal to be disentangled and some components to be more easily modeled and predicted compared to others. The following plots related to Web hits at one minute intervals. Shown are forecasts in the continuous line, and the original data in the dotted line. From upper left to bottom we have the first, second, third, fourth and residual scales. The forecasting method used here was an innovative dynamic recurrent connectionist method.