LiftKurven

Lift curves are a visualization used in predictive modeling and marketing analytics to evaluate the effectiveness of scoring models in identifying positive cases. In binary prediction, lift is defined as the ratio of the positive rate within a selected top portion of the ranked data to the overall positive rate in the population. A lift greater than 1 indicates enrichment of positives in that portion, while a lift below 1 indicates depletion.

Construction typically involves: sorting observations by the model’s score in descending order, dividing the ranked list into equally sized groups (often deciles or ventiles), and calculating the positive rate in each group and dividing by the overall rate to obtain the lift for that group. The resulting lift per group can be plotted to form the lift curve. A baseline of 1 corresponds to random selection.

Interpretation: A steeper lift curve above 1 implies greater predictive power. The curve is often used together with a gains chart, which shows the cumulative number of positives captured as a function of the population fraction. The area under the curve relates to the model’s ability to enrich positives.

Applications: lift curves are used to guide marketing campaigns, risk scoring, and resource allocation by identifying which segments to target. They can help compare models and select score thresholds.

Limitations: lift is sensitive to the base rate and class imbalance. The curve depends on the population and can change across datasets, so it should be used with calibration and in combination with other metrics like ROC-AUC or Qini.

See also: gains chart, ROC curve, AUC, Qini coefficient.