Accuracy & evaluation

RMSE

Also called root mean squared error

RMSE is the square root of the average squared error. Big misses hurt more than small ones. Use it when a rare large error is worse than many small ones — not because it looks more 'statistical.'

Squaring is a policy. It says a 10-unit miss is more than twice as bad as a 5-unit miss. Sometimes that matches a stockout or an SLA. Sometimes it just makes a noisy series look dramatic.

RMSE is in the same units as the target (after the root). It is still scale-dependent. Do not rank models on RMSE across different series without scaling.

If you optimize RMSE and report WAPE, you may pick a different champion than the one ops wanted. Align the training loss with the decision, or at least know they differ.