What is uplift modeling?
A typical propensity model ranks customers by how likely they are to buy. The top of that list is full of loyal customers who'd buy with or without a discount. Spending on them looks efficient in reports and changes almost nothing.
Uplift modeling ranks customers by the difference the marketing makes: their chance of buying if contacted, minus their chance if not. The top of that list is the people a campaign actually persuades.
How does uplift modeling work?
| Group | Buys if contacted | Buys if not contacted | What to do |
|---|---|---|---|
| Persuadables | Yes | No | Target them, since this is where the lift is |
| Sure things | Yes | Yes | Skip them, since the spend is wasted |
| Lost causes | No | No | Skip them |
| Sleeping dogs | No | Yes | Don't contact them, since marketing drives them away |
Nobody can observe both outcomes for one person, so uplift models learn from a randomized test where some people were contacted and some weren't. A simple approach trains one model on each group and subtracts the predictions. Results are checked with an uplift or Qini curve, which shows how much extra lift the model captures as more people are targeted.
What is an example of uplift modeling?
A retention offer that backfired
A subscription brand sends a discount to every customer its churn model flags as likely to cancel. Cancellations among those customers barely move.
An uplift model trained on a randomized test shows why. Many flagged customers were going to stay, and some long-time subscribers canceled after the offer reminded them they were paying. Targeting only the persuadable group cuts the cost of the program and lowers cancellations more than the blanket offer did.
Uplift model vs. propensity model
| Propensity model | Uplift model | |
|---|---|---|
| Predicts | Will this person act? | Will contacting them change whether they act? |
| Training data | Historical behavior | A randomized test with a holdout |
| Top of the ranked list | Likely buyers, including sure things | Persuadable people |
| Risk | Paying for sales that would have happened | Needs enough test data to be stable |
Why does uplift modeling matter?
Targeting by likelihood to buy rewards a campaign for finding customers who were going to buy anyway. Uplift modeling points the budget at the people it can change.
- Less wasted spend. Discounts and contacts stop going to people who'd act anyway.
- Fewer own goals. Sleeping dogs are left alone, so marketing stops pushing some customers away.
- A link to incrementality. The model's targets are chosen by causal lift, which is how the campaign will be judged.
How does Alli do this?
Uplift starts with knowing who a campaign is for. Alli Audience Planner builds Live Personas from billions of behaviors across streaming, social, commerce, creators, and professional networks, combined with a brand's first-party insights, and keeps them current as behavior shifts.
Teams compare estimated reach across platforms, test ideas with a persona through Ask Alli before spending, and carry the same audiences into Alli Media Planner, so planning and measurement start from one definition of the audience.
Explore Alli Audience PlannerUplift modeling FAQ
What is the difference between uplift and response modeling?
A response model predicts who will buy. An uplift model predicts who will buy because of the marketing, by comparing likely behavior with and without it.
What are the four customer types in uplift modeling?
Persuadables buy only if contacted. Sure things buy either way. Lost causes don't buy either way. Sleeping dogs buy only if they're left alone.
What data do you need for uplift modeling?
Results from a randomized test, where similar people were randomly contacted or not, plus the customer attributes the model uses to predict who responds.
What is a Qini curve?
It's a chart for judging an uplift model. It shows the extra conversions gained as more people are targeted in the model's order, compared with targeting at random.
Last updated Back to the glossary