Predicting Truck Requirements at Coupang

Coupang aims to optimize its inbound logistics by accurately predicting the number of trucks needed at its fulfillment centers. This involves minimizing wasted resources during product reception and ensuring timely deliveries. The process begins with feature extraction from historical logistics data, followed by training a machine learning model using the LightGBM algorithm, which efficiently handles categorical data and provides high predictive accuracy.

Model Training and Integration

Hyperparameters for the model are optimized using Bayesian optimization, ensuring continuous improvement with new data. The trained model is integrated into the inbound reservation system to provide real-time predictions to vendors. This system aims to balance underprediction and overprediction, ultimately reducing the number of unnecessary reservations.

Results and Future Plans

The final model achieves a 2.53% underprediction and a 5.04% overprediction rate, significantly improving upon previous metrics. This has led to a 67.9% decrease in delivery date changes due to slot shortages. Moving forward, Coupang plans to enhance the model's accuracy to cater to a broader range of products and fulfillment centers.