Lift and Gain Charts
Diagnostic charts that show how much better a ranking model does at finding positive cases than picking at random, by looking at the top-scored slice of predictions rather than the whole dataset.
A ranking model — say, one scoring customers by likelihood to respond to an offer — is often used by only acting on the highest-scored slice, not the whole population. A gain chart plots, for each top fraction of the ranked population (top 10%, top 20%, and so on), what fraction of all actual positives that slice contains. A lift chart takes the same idea and divides it by what random selection would have captured, showing how many times better than random the model does at each cutoff.
Lift and gain charts measure a model by how well it concentrates true positives into the top-scored slice of predictions, which is exactly how ranking models are actually used in practice — not by evaluating every prediction equally.
Reading the chart
If a model's top 10% of ranked customers contains 40% of all actual responders, the gain chart shows 40% cumulative gain at the 10% mark, and the lift chart shows a lift of 4x (40% divided by the 10% a random sample would capture). A perfect model would front-load all positives into the earliest slices; a useless model produces a straight diagonal line matching random selection exactly.
These charts are common in marketing response modeling and credit scoring, where the actual decision — call the top N% of customers, approve the top N% of applicants — depends on ranking quality at a specific cutoff, not on the overall accuracy across every score.
Related concepts
Practice in interviews
Further reading
- Practitioner note: lift charts in marketing and credit scoring