An unbiased GM(1,1)-based new hybrid approach for time series forecasting

Abstract

Purpose – The time series forecasting is an essential methodology which can be used for analysing time series data in order to extract meaningful statistics based on the information obtained from past and present. These modelling approaches are particularly complicated when the available resources are limited as well as anomalous. The purpose of this paper is to propose a new hybrid forecasting approach based on unbiased GM(1,1) and artificial neural network (UBGM_BPNN) to forecast time seriespatterns topredictfuture behaviours. Theempiricalinvestigation wasconducted byusing daily share prices in Colombo Stock Exchange, Sri Lanka. Design/methodology/approach – The methodology of this study is running under three main phases as follows. In the first phase, traditional grey operational mechanisms, namely, GM(1,1), unbiasedGM(1,1)andnonlineargreyBernoullimodel,areused.Inthesecondphase,thenewproposed hybrid approach, namely, UBGM_BPNN was implemented successfully for forecasting short-term predictions under high volatility. In the last stage, to pick out the most suitable model for forecasting withalimitednumberofobservations,threemodel-accuracystandardswereemployed.Theyaremean absolute deviation, mean absolute percentage error and root-mean-square error. Findings – The empirical results disclosed that the UNBG_BPNN model gives the minimum error accuracies in both training and testing stages. Furthermore, results indicated that UNBG_BPNN affords the best simulation result than other selected models. Practical implications – The authors strongly believe that this study will provide significant contributions to domestic and international policy makers as well as government to open up a new direction to develop investments in the future. Originality/value – The new proposed UBGM_BPNN hybrid forecasting methodology is better to handle incomplete, noisy, and uncertain data in both model building and ex post testing stages.