How do I read a confusion matrix for a financial classification model?

Question

AcadiFi · Accepted Answer

A confusion matrix is a 2×2 table showing predicted vs. actual class counts. Convention: rows = actual, columns = predicted.

|  | Predicted Default | Predicted No-Default |
|---|---|---|
| **Actual Default** | TP | FN |
| **Actual No-Default** | FP | TN |

Example: Quellmont Bank's 2025 corporate default model scored 8,200 borrowers. Actual defaults = 196. Results:

|  | Pred Default | Pred No-Default |
|---|---|---|
| **Actual Default (196)** | 143 (TP) | 53 (FN) |
| **Actual No-Default (8,004)** | 411 (FP) | 7,593 (TN) |

Derived metrics:
- Accuracy = (143+7593)/8200 = 94.3%
- Precision = 143/(143+411) = 25.8%
- Recall (sensitivity) = 143/(143+53) = 73.0%
- Specificity = 7593/(7593+411) = 94.9%
- F1 = 2(0.258×0.730)/(0.258+0.730) = 0.381
- False Positive Rate = 411/8004 = 5.1%

Business interpretation: model catches 73% of true defaulters (good recall), but 74% of flagged borrowers don't actually default (low precision). Credit committee must decide: reject 554 loans to catch 143 real defaults, or loosen threshold and accept more losses.

Cost analysis: if avg default loss is $1.8M and avg good loan profit is $140K, rejecting 411 good loans costs $57.5M in foregone profit to save 143 × $1.8M = $257M in defaults. Net benefit $199M justifies the false positive rate. CFA quantitative curriculum stresses that confusion matrix analysis must be tied to economic cost structure, not just statistical metrics.

How do I read a confusion matrix for a financial classification model?

Master Level II with our CFA Course

Related Questions

Practice Questions