Hierarchical Forecasting Method Based on ARIMAX and Recurrent Neural Network for Motorcycle Sales Prediction
I. Gede Surya Adi Prayoga, Suhartono Suhartono, Santi Puteri Rahayu · International Journal of Applied Mathematics & Statistics/International journal of applied mathematics and statistics · 2014
Motorcycle sales forecast is needed by motorcycle distributors to predict the market share and determine sales target. The purpose of this study is to forecast motorcycle sales in Pacitan, Indonesia.There are four variables to be forecasted, i.e., Matic, Cub, Sport and total sales.Motorcycle sales data has a hierarchical structure, where total sales is the aggregation of Matic, Cub and Sport sales. Consequently, it needs to be forecasted by using hierarchical method. The first step of hierarchical forecasting is obtaining individual forecast of each variable.In this step, we have compared ARIMAX and Recurrent Neural Network (RNN), and we get that RNN is better than ARIMAX, so the individual forecast is calculated by using RNN.The individual forecast does not have a hierarchical structure, so it must be revised.To revise it, we use three different methods, i.e., bottom-up, top-down and optimal combination. The results show that the best forecast for total and Matic sales are produced by top-down method, while the best forecast for Cub and Sport sales, respectively are produced by bottom-up and optimal combination method. Moreover, motorcycle sales are predicted to increase continuously and reach high sales in the months around the Eidal-Fitr holidays, except the Cub sales which are predicted to be stationary with seasonal pattern.