Gini Coefficient Metric¶

The gini metric calculates the Gini coefficient, measuring a model's ability to discriminate between defaults and non-defaults.

Metric Type: gini

Gini Calculation¶

The Gini coefficient is calculated as Gini = 2 × AUC - 1, where AUC is the Area Under the ROC Curve:

1.0: Perfect discrimination
0.0: No discrimination (random)
-1.0: Perfect inverse discrimination

The Gini coefficient is a normalized version of AUC that ranges from -1 to 1, making it easier to interpret than AUC alone.

Configuration Fields¶

Record-Level Data Format¶

For individual loan/account records:

metrics:
  discrimination_analysis:
    metric_type: "gini"
    config:
      name: ["model_discrimination"]
      data_format: "record_level"
      prob_def: "model_score" # Column with predicted probabilities (0.0-1.0)
      default: "default_flag" # Column with default indicators (0/1 or boolean)
      segment: [["product_type"]] # Optional: segmentation columns
      dataset: "loan_portfolio"

Summary-Level Data Format¶

For pre-aggregated risk-ordered data:

metrics:
  summary_discrimination:
    metric_type: "gini"
    config:
      name: ["risk_grade_gini"]
      data_format: "summary_level"
      mean_pd: "avg_probability" # Column with mean probabilities (for ordering)
      defaults: "default_count" # Column with default counts
      volume: "total_count" # Column with total observation counts
      segment: [["model_version"]] # Optional: segmentation columns
      dataset: "risk_grade_summary"

Required Fields by Format¶

Record-Level Required¶

name: Metric name(s)
data_format: Must be "record_level"
prob_def: Probability column name
default: Default indicator column name
dataset: Dataset reference

Summary-Level Required¶

name: Metric name(s)
data_format: Must be "summary_level"
mean_pd: Mean probability column name (used for risk ordering)
defaults: Default count column name
volume: Volume count column name
dataset: Dataset reference

Optional Fields¶

segment: List of column names for grouping

Output Columns¶

The metric produces the following output columns:

group_key: Segmentation group identifier (struct of segment values)
volume: Total number of observations
defaults: Total number of defaults
odr: Observed Default Rate (Defaults/Volume)
pd: Mean Predicted Default probability
gini: Calculated Gini coefficient (-1.0 to 1.0)
curve_data: Plot data for Gini/Lorenz curves (struct array with pct_observations and pct_defaults)

Curve Data for Visualization¶

The curve_data column contains plot points for creating Gini coefficient/Lorenz curve visualizations in BI tools like Power BI. Each data point contains:

pct_observations: Cumulative percentage of observations (x-axis)
pct_defaults: Cumulative percentage of defaults (y-axis)

The data starts at (0,0) and ends at (100,100), with points ordered by descending risk score. This allows direct plotting of:

Lorenz Curve: Shows concentration of defaults
Gini Curve: Visual representation of model discrimination
45-degree reference line: Represents random model performance

Example Usage in BI Tools:

Extract curve data from the result
Create scatter plot with pct_observations as x-axis, pct_defaults as y-axis
Add 45-degree reference line from (0,0) to (100,100)
Area between curves represents discrimination power

Fan-out Examples¶

Multiple Discrimination Tests¶

metrics:
  gini_analysis:
    metric_type: "gini"
    config:
      name: ["portfolio_gini", "product_gini", "region_gini", "vintage_gini"]
      segment: [null, ["product_type"], ["region"], ["origination_year"]]
      data_format: "record_level"
      prob_def: "risk_score"
      default: "default_indicator"
      dataset: "model_validation_data"

This creates four Gini metrics:

Overall portfolio discrimination
Discrimination by product type
Discrimination by region
Discrimination by origination vintage

Model Comparison¶

metrics:
  model_gini_comparison:
    metric_type: "gini"
    config:
      name: ["champion_model"]
      segment: [null]
      data_format: "record_level"
      prob_def: "champion_score"
      default: "default_flag"
      dataset: "ab_test_data"

  # Separate config for challenger model
  challenger_gini:
    metric_type: "gini"
    config:
      name: ["challenger_model_score"]
      data_format: "record_level"
      prob_def: "challenger_score"
      default: "default_flag"
      dataset: "ab_test_data"

Summary-Level Analysis¶

metrics:
  risk_grade_analysis:
    metric_type: "gini"
    config:
      name: ["overall_grade_gini", "product_grade_gini"]
      segment: [null, ["product_type"]]
      data_format: "summary_level"
      mean_pd: "grade_mean_pd"
      defaults: "grade_defaults"
      volume: "grade_volume"
      dataset: "risk_grade_stats"

Combined Discrimination Analysis¶

metrics:
  # Calculate both AUC and Gini for comprehensive analysis
  auc_metrics:
    metric_type: "auc"
    config:
      name: ["model_auc"]
      data_format: "record_level"
      prob_def: "risk_score"
      default: "default_flag"
      dataset: "validation_data"

  gini_metrics:
    metric_type: "gini"
    config:
      name: ["model_gini"]
      data_format: "record_level"
      prob_def: "risk_score"
      default: "default_flag"
      dataset: "validation_data"

Data Requirements¶

Record-Level Data¶

One row per loan/account
Probability column: numeric values between 0.0 and 1.0
Default column: binary values (0/1 or boolean)
Sufficient data points for meaningful Gini calculation (minimum ~20 observations recommended)

Summary-Level Data¶

One row per risk grade or aggregated group
Data should be ordered by risk (mean_pd column used for ordering)
Mean probabilities: numeric values between 0.0 and 1.0
Default counts: positive numbers or None (negative values not allowed)
Volume counts: positive numbers or None (negative values not allowed)
At least 2 risk grades with both defaults and non-defaults

Gini Interpretation¶

Gini > 0.6: Excellent discrimination (equivalent to AUC > 0.8)
Gini > 0.4: Good discrimination (equivalent to AUC > 0.7)
Gini > 0.2: Acceptable discrimination (equivalent to AUC > 0.6)
Gini ≤ 0.2: Poor discrimination (equivalent to AUC ≤ 0.6)
Gini = 0.0: No discrimination (random model, AUC = 0.5)
Gini < 0.0: Inverse discrimination (model predicts opposite of reality)

Relationship to AUC¶

The Gini coefficient is directly related to AUC:

Gini = 2 × AUC - 1
AUC = (Gini + 1) / 2

Why use Gini instead of AUC?

Centered around zero: Makes it easier to interpret no discrimination (0.0)
Industry standard: Widely used in credit risk modeling
Symmetric range: [-1, 1] range is more intuitive than AUC's [0, 1]
Regulatory reporting: Often required for credit risk model validation

Important Notes¶

Data Quality: Remove accounts with missing or invalid probability scores
Sample Size: Larger samples provide more reliable Gini estimates
Population Stability: Gini can vary across different populations or time periods
Risk Ordering: For summary-level data, ensure groups are properly risk-ordered
Equivalent to AUC: Gini and AUC provide identical ranking power information
Calculation Method: Uses sklearn's roc_auc_score internally for consistency

Edge Cases¶

All defaults or no defaults: Gini returns None (undefined)
Perfect separation: Gini = 1.0
Completely random: Gini = 0.0
Perfect inverse: Gini = -1.0
Insufficient data: Single observation returns None