A courier company predicts delivery durations in minutes. Observed labels stay between 0 and 100 and never sit near zero, and planners want an error figure they can quote as a percentage of the true duration. Which metric should the team report?
Show answer and explanation
Answer: B
Mean absolute percentage error averages the absolute residual divided by the actual value and reports it as a percentage, so a six-minute miss on a sixty-minute delivery reads as ten percent. That makes it easy for planners to compare errors across routes of different lengths. It suits a continuous target that is bounded away from zero, which is why the no-near-zero condition matters: dividing by a tiny actual inflates the ratio without limit. The other candidates are classifier metrics and say nothing about the size of a numeric error.
Why the other options are wrong
- A. Log loss scores predicted class probabilities against binary labels, so it has no meaning for a continuous duration target.
- C. AUC-ROC measures how well a classifier ranks positives above negatives across thresholds, not how far numeric predictions land from the truth.
- D. Precision is the share of positive calls that were correct, a classification ratio that cannot be computed on a continuous output.