Cluster-based prediction for product sales of E-commerce after COVID-19 pandemic

After a public emergency, predicting product sales in e-commerce can help to better understand and respond to market uncertainties and fluctuations. This can be of significant importance for business decision-making and inventory management. Therefore, in this study, we propose a novel sales forecasting model, which combines the dynamic time warping K-means clustering approach with the convolutional neural networks, Long Short-term Memory Networks, and Attention mechanism (called DKCLA) to predict E-commerce sales after the pandemic. Specifically, products with similar sales patterns are clustered. After that, the data within each cluster are used to construct a predictive model. In this case, if a new city experiences an outbreak of COVID-19 and the sales data in the early stage are obtained, the trained predictive model can be employed to predict product sales after lifting the lockdown in the city. Real sales data from a certain E-commerce platform are collected to verify the effectiveness of DKCLA. The results demonstrate that the proposed DKCLA model outperforms the other 36 benchmarks. In addition, the cluster-based prediction algorithm performs better than the non-clustered prediction algorithm in predicting product sales after the pandemic, and the number of clusters directly affects the prediction. And the learning rate and LSTM units exert great influence on the model performance.