Skip to content
ML Foundations

Metrics That Matter

Accuracy Is a Trap

7 min+65 XPStep 1/11

The 99.9% model that does nothing

One transaction in 10,000 is fraudulent. A 'model' that answers legit every single time scores 99.99% accuracy — and catches zero fraud.

This is class imbalance: when one class is rare, the majority drowns the metric. Accuracy quietly becomes 'how often is the answer the common one?' — which has nothing to do with whether your model learned anything.

warning

Before celebrating accuracy, compute the lazy baseline: 'predict the majority class every time.' If your model barely beats a classifier that can't learn, it probably didn't.