Hybrid SVD-ARIMA Method for Sales Forecasting with Sparse Data on E-Commerce Products

In online business, predictive analytics or forecasting is often used to improve performance effectiveness. One of the forecastings that play an important role in most businesses is sales forecasting. The results of sales forecasting are used to make stock planning and the right decisions for the future. Several previous studies on forecasting preferred to use available methods. Besides, there were also some studies that combined or compared several forecasting methods to produce higher accuracy. However, in the testing process, those studies were still carried out with non-sparse data. Therefore, the Hybrid method between Singular Value Decomposition (SVD) and Autoregressive Integrative Moving Average (ARIMA) is used to do sales forecasting in this study. SVD method is used to predict sparse data. The ARIMA method is then used to forecast sales based on data from the SVD method. The research results on monthly forecasting using sparse data of 40% have an average RMSE and MAE values improvement of 0.308 and 0.352, respectively. For monthly forecasts that use 50% sparse data, the average RMSE and MAE values improvement are 0.279 and 0.28, respectively. For daily forecasting using sparse data of 40%, the average RMSE and MAE values improvement are 0.021 and 0.014, respectively. For daily forecasting using 50% sparse data, the average RMSE and MAE values improvement are 0.017 and 0.009, respectively. The accuracy results show that the Hybrid SVD-ARIMA method can perform forecasts better than the ARIMA method. However, in daily forecasting, the Hybrid SVD-ARIMA method still has a high forecasting error.