Gini coefficient (of a model)
Also known as: Gini index · Gini · Accuracy ratio · Somers' D
A measure of how well a score orders risk: 0% is random ordering and 100% is perfect separation between cases that had the event and cases that did not. For models it equals 2 × AUC − 1.
Legal basis
scikit-learn — roc_auc_score (Gini = 2·AUC − 1)
Gini does not measure accuracy, it measures ordering. It answers one question: if you line the portfolio up from worst to best score, how well does that line separate the cases that had the event from the ones that did not. A Gini of 0% is a score that orders at random; 100% would be a score that places every event ahead of every clean case.
In risk work it is the same number as the area under the ROC curve, rescaled: Gini = 2 × AUC − 1. That is why a Gini of 20% is already a real gain over chance, 30% is good, and 50% is strong for a behavioral score — the reading depends on what you are predicting and in which population. The figure alone says little without the confidence interval around it.
Frequently asked questions
What is a good Gini value?
It depends on what the score predicts and on the population. For a behavioral risk score, the usual reading is: around 20% is a real gain over chance, 30% is good, and 50% is strong. The number is only conclusive alongside its confidence interval.