Understanding Gini Coefficient, AUC, and CAP

Here’s a concise explanation of Gini Coefficient, Cumulative Accuracy Profile (CAP), and AUC (Area Under the ROC Curve), along with their relationships:

1. Cumulative Accuracy Profile (CAP)

• What it is: The CAP curve (also called the Lorenz curve in credit risk modeling) evaluates the effectiveness of a classification model (e.g., credit scoring). It compares the cumulative proportion of positive outcomes (e.g., defaults) against the cumulative proportion of observations ranked by model scores.
• How it works:
  • X-axis: Cumulative % of observations (ordered by model score from riskiest to safest).
  • Y-axis: Cumulative % of actual positive cases (e.g., defaults).
• Perfect Model: A curve that reaches 100% of positives with the fewest possible observations.
• Random Model: A diagonal line (45°).

2. AUC (Area Under the ROC Curve)

• What it is: The AUC measures the discriminative power of a binary classifier (e.g., default vs. non-default). It’s derived from the ROC curve, which plots:
  • X-axis: False Positive Rate (FPR).
  • Y-axis: True Positive Rate (TPR).
• Interpretation:
  • AUC = 1: Perfect classifier.
  • AUC = 0.5: Random classifier.
  • Higher AUC = Better ranking ability.
• AUC is mathematically equivalent to the concordance probability (C-statistic). Concordance (or C-statistic) measures how well a binary classification model ranks predictions. It answers: “What is the probability that a randomly chosen positive instance is ranked higher than a randomly chosen negative instance?”
• Range: 0 to 1 (higher = better ranking).
• Perfect concordance (1.0): Every positive is ranked above every negative.
• Random concordance (0.5): No ranking power (like flipping a coin).
• Proof:
  • AUC = Probability that a random positive (class=1) has a higher predicted score than a random negative (class=0).
  • This is exactly the definition of concordance!

• Why Use AUC?
  • Works well with imbalanced data (unlike accuracy).
  • Measures ranking ability (how well the model orders predictions).
  • Threshold-independent (evaluates performance across all thresholds).

• Example: AUC = 0.85 means the model has an 85% chance of correctly ranking a random positive case higher than a random negative case.

3. Gini Coefficient

• What it is: A metric derived from the CAP curve or ROC curve, quantifying inequality in prediction power.
• Range: 0 (random) to 1 (perfect).
• Calculation:

• From CAP :

• From AUC: 

Gini = 2 × AUC−1

• Interpretation:
  • Gini < 0.2: Poor model.
  • Gini 0.2-0.4: Moderate model.
  • Gini > 0.4: Strong model.
  • Gini > 0.6: Excellent (rare in credit scoring).

Key Relationships

1. AUC vs. Gini:
   • Gini=2×AUC−1
   • Example: AUC = 0.8 → Gini = 0.6.
2. CAP vs. ROC:
   • CAP focuses on actual positives (e.g., defaults), while ROC considers both TPR and FPR.
   • Both can be used to derive Gini.
3. Use Cases:
   • Credit Risk: CAP/Gini are more intuitive (directly shows default capture).
   • General ML: AUC is more common (balanced view of TPR/FPR).

---

Summary Table

Metric	Source Curve	Range	Interpretation
AUC	ROC Curve	0.5 to 1	Ranking power of the model.
Gini	CAP or ROC	0 to 1	Inequality in prediction (scaled AUC).
CAP Curve	Lorenz-like curve	Visual	Shows model’s default capture rate.

I have written similar articles which may be helpful for you. Please check them out

• Evaluating Predictive Model Performance: Key Metrics
• How to Test Performance of the Linear Regression Models