Forecasting Short Time Series using Rolling Grey Bayesian Framework

Abstract

Agriculture is the backbone of Indian economy; it plays a crucial role in the overall socio-economic development of the nation. To address the needs of a growing global population and poverty eradication, appropriate policy measures in agriculture is required for economic planning and decision making. Statistical forecasting models as well as artificial intelligence approaches are generally used to predict the future pattern. Several forecasting models viz., auto regressive integrated moving average (ARIMA), auto regressive conditional heteroscedastic or generalized auto regressive conditional heteroscedastic (ARCH or GARCH), artificial neural network (ANN), support vector machine (SVM) are largely being used for forecasting production behavior of different agricultural crops using mainly time series data. These modelling techniques perform well with large amount of data or in case of situations where sufficient information is available for the crops under consideration. In real life scenario, most often the situation involves time series with comparatively short length and it is almost impossible to find parallel methods for modelling and forecasting with short time series data. Among many different available forecasting methods, grey system theory based model, rolling grey models, Bayesian (grey Bayesian models), approaches all assumed to be alternative choice for dealing with short time series data for prediction and limited number of such studies have been reported in the context of Indian agriculture. In this study, an approach has been considered by combining three different methods viz. rolling mechanism, grey model and Bayesian estimation techniques. The proposed approach is also compared with individual forecasting method. The study has been illustrated with the yearly production data of rice, wheat, maize, total oilseeds and total pulses production of Uttarakhand state and demonstrated that rolling grey Bayesian models are found to be better compared to other competitive models in accurately forecasting the production behaviors in all cases.