To measure the performance of a statistical model, namely its abilities to generalize, a contingency table is usually built. In the case of payment fraud, since payments either are fraudulent or are not, and since the statistical algorithm predicts whether the data point is fraud, there are four possible cases.
Depending on the combination of the real class and the prediction, the contingency table counts the number of true positives and true negatives (accurate predictions), and the number of false positives and negatives (errors).
The table below describes a contingency table summarizing the number of cases accurately and erroneously classified.
|Predicted Fraud||Predicted Not Fraud|
|Real Fraud||True Positive (TP)||False negative (FN)|
|Real Not Fraud||False positive (FP)||True negative (TN)|