When one class is rare, a model can look accurate while being useless.
One class has far more examples than the other.
A "model" that always answers not fraud:
accuracy 99.8%
Frauds it catches:
0%