Uplift

An online shop sends out discount vouchers and afterward measures: those who received the voucher bought more often. Sounds like success, but it isn’t proof yet. Perhaps many of these people would have bought anyway, without any discount at all. Uplift is exactly the difference that remains once you subtract that out: the additional effect that a measure truly triggers in a specific person. The term originates from marketing, but today it is used everywhere one wants to separate the effect of an intervention from mere prediction. So the question isn’t “Who buys?” but “For whom does my action change something?”.

Why discounts go to the wrong people

Classic prediction models sort customers by purchase probability. The most loyal regular customers end up at the top. They’re the ones who then get the voucher — and redeem it for something they would have bought anyway. The shop gives away margin without generating a single additional sale. In industry jargon, such customers are called a “sure thing”: high purchase probability, but uplift close to zero.

Even more unpleasant are the so-called sleeping dogs. These are people who are reminded by a promotional email that they have a subscription — and promptly cancel it. For them, the uplift is negative; the measure causes harm. A normal model can’t even detect this group, because it only considers a single state.

The economic difference is considerable. If a mail-order retailer prints and sends out a million catalogs, every contact costs money. Sending a campaign only to the high-uplift group often halves the costs while achieving the same revenue. That’s exactly why uplift is a standard topic in data teams at retail, banking, and telecommunications companies.

Two worlds that never coexist

The fundamental problem is logical in nature. One would need to know how one and the same person behaves both with a voucher and without one. Both at once is impossible. The second world is always unobservable; experts speak of the fundamental problem of causal inference. Uplift can therefore never be measured directly on a single individual.

The way out is an experiment with a control group. Customers are randomly split into two halves. One half receives the measure, the other doesn’t. Because the assignment was random, both groups are comparable on average. The difference in purchase rates is then the average uplift.

A model goes one step further. It learns from this experimental data which features correlate with high uplift — age, last purchase, basket size. A simple approach trains two separate models, one per group, and subtracts their predictions from each other. More modern methods such as uplift trees optimize directly for the difference. What remains crucial: without a clean control group in the training data, no uplift model works.

From vouchers to drug trials

Uplift is most commonly encountered in the marketing departments of large online platforms. Whoever decides which users get a push notification or a free month calculates in terms of uplift rather than click probability. Banks also use it to address only the customers who can actually be influenced during waves of cancellations.

The same way of thinking underlies every approval study in medicine. One group receives the active substance, another a placebo. Only the difference counts as the effect. In policy research, this principle is called impact evaluation, for instance when asking whether a funding program actually created jobs.

In news articles, the term sometimes appears in a weaker sense as well, such as an “uplift to the revenue forecast”. In that case, it simply means an increase. With data products, a closer look is worthwhile: if a provider advertises with “uplift”, one should ask against which control group it was measured. Without that information, the number is worth little.