The day when Drivella became more selective and stopped always swiping right — A colloquial way to explain the lift value as model assessment measure
Explaining the Lift Value can sometimes be tricky
In predictive modeling, especially for marketing and target selections the “lift value” is a frequently used measure to describe the predictive power of a model.
Sometimes it is however tricky to explain the idea and the logic behind the lift values. In my lectures for my sales and marketing students I use the following story that allows to create an easy understanding for the lift value.
Let’s Meet Drivella
In this story we meet a young lady named Drivella. She’s a big fan of online dating, and in her first half-year, she pursues a very simple approach: she accepts every invite she gets.
Over six months, she goes on 96 dates. And for each date she documents whether she met a reasonable partner. At the end of this period, she finds that she met 16 reasonable partners. That’s a success rate of about 16.7%.
Changing the Strategy
In the next half-year, Drivella changes her strategy. She becomes more selective with the acceptance of dates.
She is observing more carefully what she sees in the pictures (“is there something confusing in the background”) and she reads in the brief introductions to each person in more detail.
With this approach, she only accepts 72 dates. Looking back after she finds that she met 18 reasonable partners — raising her success rate to 25% (18 out of 72).
What Does This Have to Do With Predictive Modeling?
In predictive modeling, we often want to know: How much better is my model at selecting the target customers compared to random selection? This is where the lift value comes in.
In the example of Drivella we have 2 approaches (or models): the random model (where every invite got accepted) and Drivella’s “rules based logic”
- Random selection (first half-year): 16.7% success rate.
- Rule-based selection (second half-year): 25% success rate.
The lift value is simply the ratio of these two rates:
Lift = 25 % / 16.7 % = 1.5
This means Drivella’s new selection criteria (her “model”) helped her find reasonable partners 1.5 times more often than if she had just accepted every invite randomly.
Lift tells us how much better our model is compared to a random selection.
You can also say: The lift value is the multiplicative Factor how much your predictive model is better than a random selection.
The Baseline Model
The random selection is often also named as the “baseline” model. You can also consider this as the benchmark for the most stupid model (that has no predictive power) it will perform as good as the “baseline” random selection model.
Captured Response
Another way to compare models in predictive modeling is the captured response. In a marketing example this is the percentage of responders that you can capture with the top 10% (top 20% …) of your customers base.
- If you had a random model you would capture ~10% of the responders with a selection of 10% of the customers.
- If your models has predictive power and allows to prioritize the responders in the selection you would see captured response like 15%, 20%, … with a selection of 10% of the customers.
In Drivella’s case, if she wants to meet 10 reasonable partners:
- With random selection, she’d need about 60 dates
- With her selective logic, she’d only need 40 dates
Wrapping Up
So, the next time you hear about “lift value” in predictive modeling, think of Drivella’s dating journey. By being more selective — using a model — she improved her chances and efficiency. That’s exactly what a good predictive model should do: help you find what you’re looking for, faster and more reliably than random guessing.
