Act 2 · Foundation · 9 minutes

Every prediction has two stories.

What the model predicted—and what actually happened. Cross those two questions and four outcomes appear.

Build it, do not memorize it

  1. Choose the class we care about: fraud = positive.
  2. Ask: was the prediction positive or negative?
  3. Ask: was it correct (true) or wrong (false)?
True / false means correct / incorrect. Positive / negative means the model’s predicted class.

Audience check

The transaction was fraud, but the model said legitimate. What is it?

False Negative. The model predicted negative, and that prediction was false.

The 2 × 2 error map

Predicted positive
(fraud)
Predicted negative
(legitimate)
Actually positive
(fraud)
TPFraud caught
FNFraud missed
Actually negative
(legitimate)
FPFalse alarm
TNCorrect clearance

Rows describe reality; columns describe the model decision. Always state which class is “positive.”

Hands-on metric calculator

Adjust the four counts. Notice that the same accuracy can hide very different operational outcomes.

8
12
2
78
—Accuracy
—Precision
—Recall
—F1 score

Accuracy reconstructed

Accuracy = (TP + TN) ÷ (TP + TN + FP + FN)

Now we can see what accuracy merges together: both kinds of correct prediction. Next, we will deliberately focus on the positive class.

← Accuracy trapNext: Precision, recall & F1 →