Optimizing the Inbound Process with a Machine Learning Model

To use inbound slots at fulfillment centers more efficiently, Coupang built a machine learning model that automatically predicts the number of trucks vendors need when making reservations. Accurate demand forecasting is key, since underprediction causes inbound delays and overprediction wastes limited docks and slots.

Model development proceeded in three stages.

• Based on over 2 years of accumulated inbound request and logistics data, about 800,000 training data points were created, and meaningful features were extracted through interviews with field logistics staff and EDA.
• LightGBM was used for the model, because it trains fast even on large-scale data and performed well when there were many categorical features.
• Hyperparameters were automatically searched using Bayesian optimization, and this process was continuously updated on a monthly basis with new data.

On the operational side, the model was deployed on Amazon SageMaker, and integrated so that the inbound reservation system calls the endpoint upon a reservation request and immediately shows the prediction result. This allows the appropriate number of trucks to be guided directly within the reservation system.

The key challenge is balancing underprediction and overprediction. Predicting too few increases inbound delays, while predicting too many increases unnecessary slot reservations. So rather than simply minimizing error, the focus was on meeting prediction standards that also consider the impact on logistics operations.