Most campaign targeting asks the wrong question. A typical propensity model ranks customers by how likely they are to buy, and the campaign goes to the top of that list. But customers who are very likely to buy will often buy anyway. Sending them a ₹150 discount does not create a sale; it gives away margin on a sale you already had. The question that matters for spending money on a campaign is different: whose behaviour will this offer change? Uplift modelling is built to answer that.

Four kinds of customer

Thinking about how each customer would respond with and without the offer gives four groups, a framing widely used in the uplift literature:

  • Persuadables buy only if they receive the offer. These are the customers worth paying for.
  • Sure things buy either way. Targeting them wastes the discount.
  • Lost causes do not buy either way. Targeting them wastes contact cost.
  • Sleeping dogs are less likely to buy, or more likely to leave, because they were contacted. A retention call that reminds a customer their contract is ending, or an aggressive offer that annoys them, can trigger exactly the churn it was meant to prevent.

You can never observe which group an individual belongs to, because each person either gets the offer or does not. But you can estimate, for groups of similar customers, the difference in purchase rate between treated and untreated. That difference is the uplift.

A worked comparison

An online fashion retailer runs a pilot: within each of three customer segments, a random half receive a ₹150 voucher and the other half receive nothing. A purchase is worth ₹600 in contribution before any discount.

SegmentCustomersBuy rate with voucherBuy rate withoutUpliftProfit impact of targeting whole segment
A: recent frequent buyers20,00030%28%+2 pts−₹6.6 lakh
B: lapsed for 3–6 months20,00012%4%+8 pts+₹6.0 lakh
C: recent complainants10,0005%7%−2 pts−₹1.95 lakh

The profit column compares two scenarios for each whole segment. Targeted: buyers bring ₹600 minus the ₹150 voucher, so ₹450 each. Untargeted: buyers bring the full ₹600. For segment A, that is 6,000 buyers × ₹450 against 5,600 buyers × ₹600, a loss of ₹6.6 lakh. The voucher bought 400 extra sales but gave a discount to 5,600 people who would have paid full price.

A traditional response model would rank segment A first, since it has the highest buy rate. An uplift view ranks B first, the only segment where the campaign makes money, and flags C as one to avoid entirely: in this illustration, the voucher lowered purchases among recent complainants, perhaps because it reminded them of a bad experience.

How uplift models are built

Uplift modelling needs data from a randomised treatment and control group. Without that, you cannot separate the offer’s effect from the fact that marketing usually targets customers who were already promising. The usual approaches:

  • Two-model approach (often called a T-learner): train one model predicting purchase among treated customers and another among control customers. Each customer’s uplift is the difference between the two predictions. Simple, but the errors of two models can swamp a small true difference.
  • Single model with a treatment flag (S-learner): include “received offer” as a feature, then predict each customer twice, with the flag on and off.
  • Class transformation: with a 50/50 random split, a clever relabelling turns the problem into a single classification whose output relates directly to uplift.
  • Direct methods: uplift trees and causal forests split customers to maximise differences in treatment effect, rather than differences in outcome.

Evaluating without individual ground truth

Since true uplift is never observed per customer, accuracy cannot be measured the usual way. Instead, sort customers by predicted uplift, split them into deciles, and in each decile compare the actual buy rate of the treated and control customers there. A good model shows large real uplift in the top deciles, falling towards zero or negative at the bottom. Cumulative versions of this chart are called Qini or uplift curves. A model whose deciles all show similar uplift has learned nothing useful, however good its ordinary accuracy.

Practical difficulties

Uplift signals are small: the difference between 12% and 4% is large by campaign standards, and many real uplifts are one or two points. Detecting them reliably needs big samples, which means sacrificing revenue to a control group. Business teams often resist holding customers back, and that resistance is the most common reason uplift projects stall.

Uplift also decays. Customers who were persuadable last quarter may not be this quarter, especially if they have learned to wait for vouchers. Keep a permanent random control group in every campaign so the model can be refreshed and so the business always has an honest measure of incremental impact. Even without a full uplift model, that habit alone stops many campaigns from claiming credit for sales they did not create.

Written by the RP Strativa Team. The retailer, segments and figures are illustrative.

Leave a Reply

Trending