Demand forecasting models for planning product purchases for retailers: a case study of superstore retailers

This research aims to develop appropriate forecasting models for each product using time series forecasting methods and artificial neural networks, focusing on models with the lowest errors; weekly sales data of products in a superstore over the past four years, from January 2014 to December 2017, was used, specifically targeting Class A products, totaling 210 weeks, and the data was divided into three parts: training Set comprising 168 weeks (80%), validation Set comprising 21 weeks (10%), and test Set comprising 21 weeks (10%), then tested with the test Set, which consisted of new, unseen data; and since the data used for analysis was obtained from a source used for data visualization, it was impossible to set standard criteria for the analysis, leading to a comparison of performance and calculation of the percentage improvement of the ANN method relative to the Holt-Winters method for each dataset, finding that the ANN forecasting model is generally more effective, with the ANN method improving MAD by 30.24% and RMSE by 25.79% for phone data compared to the Holt-Winters method, whereas the Holt-Winters method remained superior for chair data, with MAD increasing by 69.72% and RMSE by 75.11%, and for storage data, with MAD increasing by 10.05% and RMSE by 17.76%, while the ANN method demonstrated improvements for table data, with MAD increasing by 21.10% and RMSE by 27.52%, ultimately highlighting that the ANN method is more effective for phone and table data, whereas the Holt-Winters method proves to be more suitable for chair and storage data, emphasizing the necessity of selecting appropriate forecasting models based on the specific characteristics and requirements of each dataset.