Development and performance evaluation of nonlinear and robust adaptive models for prediction of number of customers of mobile phone service providers

Abstract

To devise the marketing strategy for achieving business growth there is a potential need to efficiently forecast the number of customers of mobile phone service providers. Most of the existing forecasting models are either linear or employs squared error cost function for updating weights of the adaptive models. As a result these models fail to predict satisfactorily when the input attributes are nonlinearly related to predicted output or the data are contaminated with outliers. To alleviate these limitations this paper proposes efficient nonlinear as well as robust forecasting models for predicting the number of mobile phone customers of service providers. They employ extracted parameters such as mean and variance from the past data as inputs and a simple learning algorithm based on minimisation of a robust norm of errors rather than squared error term. The desired nonlinearity to the model is introduced by sine-cosine expansions of the input features. Assessment of prediction performance of the proposed model through exhaustive simulation study is found to be much superior compared to conventional prediction models. When outliers are present in the data the robust model outperforms the existing method of prediction.