I have a model which has a lower loss value than the baseline model I'm comparing to, but also slightly lower accuracy. The model was trained to classify time series into one of two classes. In general, which is a better indicator of performance and why? I have seen conflicting answers on this and I'm still confused.
Asked
Active
Viewed 8 times
0
-
1If your loss is a proper scoring rule then the loss is a better measure of performance than accuracy since accuracy can be optimized by very poor models (e.g. guessing most prevalent class) – Demetri Pananos Apr 19 '21 at 02:17
-
Please don't just repost the same question. Instead, edit the original question to clarify. – Stephan Kolassa Apr 19 '21 at 06:00