1) Meaning

The Qini Curve is a visual evaluation tool used in uplift modeling (a.k.a. causal or incremental response modeling).

  • It shows the incremental benefit (uplift) achieved when targeting customers ranked by predicted uplift score.
  • Similar to how an ROC curve evaluates classification models, the Qini curve evaluates how well an uplift model identifies Persuadables.

The higher and more “bowed upward” the Qini curve, the better the model at focusing on customers who truly respond positively to treatment.

2) Core Idea

  • Customers are ranked by uplift score (from highest → lowest).
  • You incrementally target top x% of customers.
  • For each portion, you calculate incremental responses (treatment group – control group).
  • The cumulative incremental response is plotted.

3) Components

  • X-axis: Proportion of population targeted (0% → 100%).
  • Y-axis: Cumulative incremental responses (uplift).
  • Random targeting line (baseline): shows expected uplift if you targeted randomly.
  • Model curve (Qini curve): shows how much extra uplift your model achieves compared to random.

4) Example

Suppose we run a marketing campaign with a treatment group and control group:

  • Rank customers by uplift score.
  • Target top 20%:
    • Treatment purchase rate = 18%
    • Control purchase rate = 12%
    • Incremental uplift = +6% × number of customers targeted.
  • Repeat for 40%, 60%, etc.

When plotted:

  • The Qini curve rises steeply if the model effectively finds Persuadables.
  • A flat or near-random line indicates poor model performance.

5) Qini Coefficient

  • A summary metric derived from the Qini curve.
  • Similar to AUC for ROC.
  • Defined as the area between the model’s Qini curve and the random baseline.
  • Higher coefficient → better targeting power.

6) Why It Matters

  • Evaluates causal models properly (unlike accuracy or AUC, which only measure predictive power).
  • Business impact focus: shows how many incremental conversions or revenue a campaign generates.
  • Model comparison: lets you compare uplift models (random forest, logistic regression, uplift trees, etc.).

Bottom line:
The Qini Curve is the key evaluation plot for uplift models, showing how effectively the model identifies customers who change behavior because of the treatment. Its associated Qini Coefficient quantifies the incremental impact above random targeting.