A food-delivery app trains a model to predict how many minutes each order will take to arrive. On a validation set, the current model's errors fall into three clear groups:
| group | share of orders | size of the error | what is behind it |
|---|---|---|---|
| everyday orders | to minutes | ordinary noise | |
| storm days | to minutes | real delays: these orders genuinely arrived late, customers complained, and the team wants the model to predict them well | |
| dead phones | to minutes | broken labels: the courier's phone died and logged a nonsense arrival time hours later |
The team switches to Huber loss. For an error smaller than delta it curves upward like MSE; for an error larger than delta it rises in a straight line like MAE.
Which delta fits this data best?
Select all that apply.