Enhancing the Forecasting Performance with the Hybrid Machine Learning Techniques for Online Shopping Channels: A Case Study of Electronic Products in Thailand

This study proposes novel hybrid forecasting models, which are the combination between Long-Short Term Memory (LSTM) recurrent neural network and traditional models. The objective is to investigate the trend of customer behavior on electronic products and reduce the demand variation that will be impact the total cost in the supply chain. In order to prove the performance of our proposed hybrid models, a Thai company case study is tested with four online shopping electronic products. Accuracy and demand variation are used to compare the effectiveness of our hybrid forecasting model. For the accuracy, some performance indicator tools, such as MAE, RMSE, and MASE, show that the hybrid models provide the lowest error among these indicators with three products, while the single LSTM provide the lowest error with only one product. For the demand variation, all forecasted demand provide the coefficient variation (CV) score less than 0.25, which mean that all products are less variation, and can be used to calculate the forecast stock levels in the company.