Modeling and forecasting of hourly transactions on a WWW and proxy cache server

A modeling approach is introduced in this paper to capture the real characteristics of the WWW traffic, using a combination of Box-Jenkins models with seasonality removal. Fourier analysis and standardization are used to render seasonal data into non-seasonal data. However, seasonality cannot be completely removed by this approach, so both seasonal and non-seasonal models are employed and the best models are selected using posterior possibility criterion (PPC). In order to improve the forecasting accuracy, forecast values obtained by the developed models are adjusted by the difference percentage of the day of the week and by using this percentage adjusted values in place of the actual data, which are not available at the time of forecasting. With this adjustment, the developed models can provide very accurate forecast values, except for the case where the higher level of hierarchical proxy is shared by child proxies located in a different time zone.