Wireless Data Traffic Estimation Using a State-Space Model

A new forecasting technique called the extended structural model (ESM) is presented. This technique is derived from the basic structural model (BSM) by the introduction of extra parameters that were assumed to be I in the BSM. The ESM is constructed from the training sequence using standard Kalman filter recursions, and then, the extra parameters are estimated to minimize the mean absolute percentage error (MAPE) of the validation sequence. The model is evaluated by the prediction of the total number of minutes of wireless airtime per month on the Bell Canada network. The ESM shows an improvement in the MAPE of the test sequence over the BSM, seasonal autoregressive integrated moving average (ARIMA), and generalized random walk models on the series considered in this paper. The improved prediction can significantly reduce the cost for wireless service providers who need to accurately predict future wireless spectrum requirements.