Data-driven lead time forecasting and promising for an online retail platform

Customized product fulfillment grapples with challenges in lead time prediction and promising due to the limited access to seller information on online retail platforms when selling customized products. Our work presents a data-driven analytics study on the challenges of lead time forecasting and commitment for an online retail platform selling customized products in real time. We first employ forecasting models common for distribution prediction: quantile regression forests, regression trees, and quantile regression. Utilizing these models, we harness the complex interrelations between lead time and related operational predictors to produce distribution forecasts. Tested on real-world data from our industry partner Babyonline Dress (BD), an online retail platform implementing the fully managed model and selling customized dresses, the quantile regression forests model delivers superior predictive performance. Subsequently, we proceed with lead time promise optimization by proposing a cost-minimization decision rule considering asymmetric misclassification costs, i.e., different penalties for late and early fulfillment. Numerical results based on the BD’s transaction data show that our proposed data-driven lead time forecasting and promising approaches could have helped the online retail platform reduce the expected misclassification cost by at least 27.04%. Additionally, our simulation study suggests that this new approach can achieve an 8.43% improvement in sales while keeping the order delay at a level similar to that of BD’s current lead time promise policy. Our research shows how online retail platforms of customized products can strategically set promised lead times to increase sales and improve customer satisfaction by estimating the lead time distribution more precisely.