The trap
Accuracy is the fraction of predictions you got right. It feels intuitive, so beginners reach for it first. The problem appears when classes are imbalanced.
A worked example
Suppose 1 in 1000 transactions is fraud. A model that always predicts not fraud is right 999 times out of 1000.
- Accuracy is 99.9 percent
- It catches zero fraud
- It is completely worthless for the actual goal
This is the accuracy paradox: the lazy majority class predictor wins on accuracy while failing the task.
What to do instead
- Look at the confusion matrix, not a single number
- Track precision and recall for the minority class
- Report a balanced metric like balanced accuracy or F1
- Compare against a majority class baseline so a high number means something
The lesson is not that accuracy is wrong, but that it answers the wrong question when the classes are skewed. Always state the base rate alongside any accuracy figure.
Key idea
On imbalanced data a high accuracy can mean nothing. Always compare to the majority baseline and report class aware metrics.