SummaryNetwork traffic prediction is a fundamental tool to harness several management tasks, such as monitoring and managing network traffic. Online traffic prediction is usually performed based on large sets of historical data used in training algorithms, for example, to determine the size of static windows to bound the amount of traffic under consideration. However, using large sets of historical data may not be suitable in highly volatile environments, such as cloud computing, where the coupling between time series observations decreases rapidly with time. To fill this gap, this work presents a dynamic window size algorithm for traffic prediction that contains a methodology to optimize a threshold parameteralphathat affects both the prediction and computational cost of our scheme. Thealphaparameter defines the minimum data traffic variability needed to justify dynamic window size changes. Thus, with the optimization of this parameter, the number of operations of the dynamic window size algorithm decreases significantly. We evaluate thealphaestimation methodology against several prediction models by assessing the normalized mean square error and mean absolute percent error of predicted values over observed values from two real cloud computing datasets, collected by monitoring the utilization of Dropbox, and a data center dataset including traffic from several common cloud computing services. Copyright © 2016 John Wiley & Sons, Ltd.