Question

A logistics company wants a model that predicts delivery times for shipments. Rather than a single point estimate, the operations team needs an upper bound they can commit to customers, specifically the value below which 90% of actual delivery times will fall. The data scientist trains a gradient boosting model in SageMaker. Which approach produces the required estimate directly from a single trained model?

Accepted Answer

Train the model with a quantile (pinball) loss objective set to the 0.9 quantile

Answer

Train with mean squared error and report the predicted value plus two standard deviations

Answer

Train with mean absolute error, which minimizes the median and represents the 90th percentile

Answer

Train with log-cosh loss and multiply the prediction by 1.9 to approximate the upper bound