Gains and Lift Analysis

Gains and lift analysis ranks cases by a predicted score and shows, for each depth of the ranked list, the share of responders captured and how much better than random targeting the score performs.

이득·향상도 분석은 예측 점수로 사례를 줄 세운 뒤, 상위 몇 %를 골랐을 때 반응자를 얼마나 잡고 무작위보다 몇 배 나은지를 보여 줍니다.

If I contact only the top 20% of customers by model score, what share of buyers will I reach?

모델 점수 상위 20% 고객에게만 연락하면 구매자의 몇 %에 닿을 수 있나요?

Cases are sorted by score and split into equal-size bins, and each bin's response rate, lift over the base rate and cumulative gain are reported along with the Kolmogorov-Smirnov statistic. When revenue and contact cost are given, the profit curve identifies the depth that maximizes profit.

When to use Gains and Lift Analysis

• Choosing how deep to go in a ranked list for a marketing campaign or audit.
• Communicating a classifier's value to non-technical audiences.
• Comparing targeting depth against contact costs.

When to use something else

• You need calibrated risks rather than ranking; use calibration assessment.
• The outcome is continuous; use regression error measures or decile mean plots.

Assumptions and requirements

• A binary response and a score where higher means more likely to respond.
• The score was built on different data from those evaluated, or results will be optimistic.
• Ties are broken by row order.

What the analysis reports

• Decile table with response rate, lift, cumulative gain and cumulative lift
• Kolmogorov-Smirnov statistic and its cutoff
• AUC
• Profit-optimal contact depth when value and cost are supplied

How to report it

Ranking cases by [score] (N = [value], base response rate = [value]%), the top 10% captured [value]% of responders (lift = [value]); the Kolmogorov-Smirnov statistic was [value].

Related methods

• ROC Curve and AUC
• Decision Curve Analysis (DCA)
• Binary Logistic Regression

Run Gains and Lift Analysis on Quorant