Causal inference lab · Uplift modelling
Uplift modelling vs churn prediction: who should get the offer?
Churn prediction ranks customers by how likely they are to leave. Uplift modelling ranks them by how much a retention offer would change that. The two lists differ: some high-risk customers leave whatever you send, and some customers stay only if left alone. This demo targets the same synthetic customers both ways and counts how many extra customers each ranking retains at a fixed budget.
These demos are illustrative. Every customer, account, region and order is synthetic, generated and analysed in your browser, and every number shown is computed from that synthetic data. They show general principles of causal analysis, not the results of any real campaign, product or project.
Nothing is uploaded. No customer data is used.
Uplift modelling / targeting
Most likely to leave. Not most worth an offer.
What this shows: with a fixed number of retention offers, ranking customers by estimated uplift retains more of them than ranking by churn risk, when the customers at highest risk are hard to move.
Set the budget
Both scores are fitted on a randomised pilot of 20,000 customers and applied to 20,000 others. Untick the box to make the offer work better the higher the risk.
Extra customers retained
20,000 synthetic customers rankedBlue: ranked by upliftAmber: ranked by churn riskGrey dashed: offers at randomWhite dashed: your budget
Synthetic data · illustrative · not results from a real business
Retention offers, win-back calls, discounts and any other action with a cost per customer. The business result is the number of outcomes the action changed, which a churn model’s accuracy does not measure.
What this experiment models
40,000 synthetic customers with two observed features, usage decline and past response to offers. Churn risk without the offer rises with usage decline. The offer’s effect grows with past response; with the box ticked it is largest at middling risk and smaller at the extremes, and unticked it grows with risk. Customers with little past response become slightly more likely to churn when contacted. 20,000 customers form a randomised pilot (offer by coin flip); in a 5 × 5 grid of segments the churn score is the churn rate without the offer and the uplift score is that rate minus the churn rate with the offer. The other 20,000 are ranked by each score, ties in simulation order. Extra customers retained counts customers who churn without the offer and stay with it, minus those who do the reverse; the simulation holds both outcomes for every customer. Fixed seed.
The principle
How it works
What the demo shows
The demo generates 40,000 synthetic customers with two observed features: recent usage decline, which drives churn risk, and past response to offers, which drives how much an offer helps. In the default population the offer works best on customers at middling risk. Those at the highest risk are harder to move, and customers who rarely respond to offers become slightly more likely to leave when contacted.
Half the customers form a randomised pilot in which a coin flip decides who gets the offer. Two scores are fitted on the pilot in 25 segments: a churn score, the churn rate without the offer, and an uplift score, that rate minus the churn rate with the offer. The other 20,000 customers are ranked by each score. Without any offer, 9,934 of them would churn.
With offers for 20% of customers, 4,000 offers, ranking by churn risk retains 336 customers who would otherwise have left. Ranking by uplift retains 935, 2.8 times as many, and offers sent at random retain 417. Per 100 offers that is 23.4 customers against 8.4. An offer for everyone retains 2,085; the uplift ranking peaks at 2,130 when 80% of customers are targeted, because the last customers on its list are the ones the offer pushes away.
Untick the box and the offer works better the higher the risk. The churn-risk ranking improves and the gap narrows; it does not close, because responsiveness still differs between customers at the same risk.
Why it matters for a decision
A churn model answers the question of who will leave. A retention budget needs to know whose decision can be changed. When the two differ, ranking by risk spends offers on customers who leave anyway and on customers who would have stayed anyway. The size of the gap in this demo is a property of the synthetic population; in real data it has to be measured.
What to check in your own data
- Do you have randomised data? Uplift can only be estimated and evaluated where the offer was given or withheld at random, at least in a pilot.
- Evaluate on held-out customers with an uplift curve: incremental outcomes against the share targeted, as in the chart. Accuracy or AUC on churn says nothing about it.
- Is the signal strong enough? Uplift is a difference of two rates, so it is noisier than either. With a small pilot, a simple churn ranking can do better. The demo uses 25 coarse segments for this reason.
- Look for segments with negative uplift. Customers the action harms are a reason to stop well before everyone is contacted.
For how uplift models are built and evaluated in a project, see causal inference consulting and churn prediction; for background, the causal inference guide.
FAQ
Common questions
What is the difference between churn prediction and uplift modelling?
Churn prediction estimates the probability that a customer leaves. Uplift modelling estimates how much an action changes that probability. The first identifies who is at risk; the second identifies who is worth acting on.
Does uplift modelling need an A/B test?
It needs data in which the action was assigned at random, or observational data with a credible adjustment for how it was assigned. The cleanest route is a randomised pilot: fit the model on one part and evaluate it on another.
What are persuadables, sure things, lost causes and sleeping dogs?
They are four response types. Persuadables stay only if they get the offer. Sure things stay either way. Lost causes leave either way. Sleeping dogs leave only if contacted. Only persuadables repay an offer. No individual’s type can be observed, so uplift models estimate the average effect for groups of similar customers.
When is a churn model enough?
When the action’s effect rises with risk and little else predicts the response, a churn ranking captures most of the gain. Unticking the box in the demo moves towards that case. A randomised pilot shows which case your customers are in.
Causal inference lab
More demos
All five demos on one page: the causal inference labCausal inference consulting →Causal inference guide →