How to Answer Metric Questions in Interviews

A better answer framework

When an interviewer asks “Precision or recall?”, avoid immediately choosing one.

Use this structure:

1. Define the prediction goal.
2. Identify the positive class.
3. Compare the cost of false positives and false negatives.
4. Check class imbalance.
5. Select the metric and threshold based on those costs.

Example

Question: Which metric would you optimize for a fraud model?

A strong answer could be:

I would first understand the cost of missing fraud versus incorrectly flagging legitimate transactions. If missed fraud is substantially more expensive, recall becomes important, but I would not maximize recall without constraint because excessive false positives can block good customers. I would inspect precision-recall trade-offs and choose an operating threshold based on business cost.

That answer demonstrates reasoning, not memorization.

What the interviewer is testing

They usually want to know whether you understand that model evaluation is connected to business decisions.

Likely follow-up

“What happens if you lower the classification threshold?”

In many binary classifiers, lowering the threshold tends to increase recall while potentially reducing precision because more cases are classified as positive.

Practice

How would your answer change for spam filtering, cancer screening, and credit-risk approval?