Gains Chart

A gains chart, also called a cumulative gains chart, is a way to judge how well a predictive model sorts a population by likelihood of some outcome — a customer responding, a lead converting, an account churning. You take the model's scores, rank everyone from most likely to least likely, and split them into ordered groups, usually deciles. Then you plot two things: on the horizontal axis, the cumulative share of the population you have targeted as you move down the ranking; on the vertical axis, the cumulative share of all the actual positive outcomes those people account for. The result is a curve that answers a practical question — if I contact the top 10%, 20%, or 30% my model flags, what fraction of the real responders do I catch?

The chart earns its value by comparison. A diagonal baseline runs corner to corner, representing random selection: target 30% of people at random and you would expect to reach roughly 30% of the responders. A useful model bows well above that line, because its high-scoring groups are dense with real positives. If the top decile contains, say, half of all responders, the curve shoots up steeply at the left before flattening. The bigger the gap between the model's curve and the random diagonal, the better the model concentrates the outcomes you care about into the groups you would contact first. A perfect model traces the highest possible path — capturing all positives as early as the true base rate allows — while a worthless one hugs the diagonal.

A gains chart is closely related to two other model-evaluation plots, and keeping them straight prevents confusion. A lift chart is the gains chart's ratio cousin: instead of showing the cumulative share of outcomes captured, it shows lift — how many times more positives a targeted group holds compared with random selection at the same depth. If the top decile captures 50% of responders when 10% would be expected by chance, that is a lift of about five. Gains and lift carry the same information from the same ranking; the gains chart reads as a cumulative share climbing toward 100%, while the lift chart reads as a multiple that starts high and falls toward one as you target everyone.

The ROC curve is a different animal, though people mix it up with the gains chart. ROC plots the true-positive rate against the false-positive rate across score thresholds, and it is aimed at classification quality independent of how common the outcome is. A gains chart is aimed at targeting and ranking — its horizontal axis is the share of the whole population you contact, which is exactly the lever a marketer pulls when deciding how deep to mail, call, or advertise. Both come from the same ranked scores, but the gains chart speaks the language of campaigns and budgets, answering how many outcomes you capture for a given reach, while ROC speaks the language of classifier accuracy. Choose the one that matches the decision at hand.

A gains chart is most useful when you have a limited budget and must decide how deep to go into a ranked list. Read it to find the point of diminishing returns: where the curve starts to flatten, each additional slice of the population brings in fewer new responders, so extending the campaign there buys little. If contacting the top three deciles captures most of the responders and the curve levels off after that, the chart is telling you to stop around there and spend the rest elsewhere. Used this way, it turns a model's scores into a concrete targeting decision — how many people to reach, and what share of the possible wins that reach buys.

The failures come from misreading or over-trusting the chart. Building it on the training data rather than a held-out sample flatters the model, so evaluate on data the model has not seen. Judging a model by a single point on the curve hides how it behaves at other depths. And a steep gains curve only matters if the outcome it ranks is the one that pays — a model that concentrates cheap, low-value responders can look great on a gains chart while adding little profit. Read the chart alongside the economics: the goal is not the prettiest curve but the targeting depth that captures the most valuable outcomes for the budget you have.

A gains chart — from the idea of the gain a ranked model gives over random selection — is a standard tool in predictive modeling and data mining for evaluating and applying scoring models.

Etymology: source.

Search interest for this term over the last five years:

View interest-over-time on Google Trends →

Areas of marketing where gains chart is a core concern: