Predicting Web Site Access: An Application of Time Series

modeling of the connections per day with time series did not give good results (the errors in some case exceeded the 200%), one of the reason of this bad performance was represented by the sharp fluctuations of the number of connections. On the other hand the results obtained when modeling the average number of connections per week were encouraging. In particular, Figure 3 and Figure 4 report respectively the one and the four steps ahead prediction errors when the average number of connections per week was preditced. The analysis of Figure 3 highlights that the biggest errors are concentrated between the ¢ ¡ ¤ £ ¦ ¥ and the § ¢ ¨ £ ¦ ¥ weeks, in coincidence with the launch of the new Master Degree and the web site restructuring. In all the other cases the prediction error can be considered acceptable: in the worst case it reach 0.23 (i.e., 23 %) and its average value In this case the obtained errors are greater than the ones obtained when the one step ahead value of the time series was predicted. It is worth noting that these errors are concerned with a medium range prediction (4 weeks) and, however, the average value of the four steps ahead prediction error is 0.28 (i.e., 28 %).