Machine learning based modeling approach to predict delivery promise time for e-commerce

Abstract

To smoothly function an online retail business, provision of reliable and fast delivery service is essential. A customer's buying decision on an ecommerce platform is discouraged due to erroneous promise timelines defined at the point of order placement. The ecommerce industry is under increasing pressure to create a data-driven framework that can accurately forecast the distribution of delivery time and set precise promise delivery times. By exploiting the relationships between delivery time and relevant predictor variables, developed tree based model, quantile regression, quantile regression forest model, LightGBM trained with quantile regression model, and LSTM deep learning model to generate conditional distribution forecasts. The models are evaluated using two methods of scoring rules namely pinball loss function and continuous ranked probability score (CRPS). LSTM model incorporated with quantile regression generated best results out of all developed, trained and tested from real world dataset obtained from ABC.com - Sri Lanka’s largest ecommerce platform, where multi-origin marketplace seller fulfillment was considered as the novelty contribution. Furthermore, it was revealed that hour of order placement and origin to destination combination impacted most on delivery time and highlighted the importance of integrating the forecasting model to decision making structure. As a limitation of this study and eligible future work, forecasting the delivery time with most recent order distribution restricting to the ongoing logistics capacity utilizations can be stated.And also consideration on incorporating real-time data and cost-sensitive approaches to improve the practicality and effectiveness of forecasting in e-commerce businesses