Loyalty glossary · 7. Data and analytics (31)
Uplift Modeling
Uplift modelling is a causal analysis technique that separates members whose behaviour a loyalty intervention will actually change from those who would have acted the same anyway, so a programme can target only the persuadable segment and measure true incremental impact.
Uplift modelling answers a question loyalty programmes usually avoid: which members would not have acted without the intervention? Most operators measure response after an offer, but that mixes cause with pre-existing behaviour. A member who redeems a 5,000 point bonus was probably already active. Without a causal split, the programme credits the bonus for behaviour it did not create.
The standard uplift segmentation divides members into four groups. Persuadables change behaviour only when treated. Sure things would have acted anyway. Lost causes will not act regardless. Sleeping dogs respond negatively, and the offer actually suppresses their activity. Targeting the first group and avoiding the last two is the entire point of the technique.
The arithmetic makes the waste visible. A programme sends a 1,000,000 point bonus to every member in a test cell and sees a 10 percent uplift in redemption, which is 100,000 incremental points. But an uplift model would have identified that only 40 percent of the cell is persuadable, so 600,000 of those points went to sure things and lost causes with zero incremental effect. That is the difference between spending to change behaviour and spending to observe it.
Uplift modelling fails when it is treated as just another propensity score. A propensity model predicts who is likely to redeem, which is not the same as who redeems because of the offer. An actuarial model does the same for aggregate liability, smoothing expected cost without saying which member changes. Accrual campaigns invite the same error: they reward earning activity without testing whether the reward caused the activity or simply followed it. Activity-based qualification has a parallel weakness, because it grants tier on observed behaviour and cannot separate members who did the behaviour to get the tier from those who would have done it anyway.
The trap is complacency after the first model. Uplift effects decay as members learn the patterns of the programme, and a segment that was persuadable for six months becomes a sure thing in the seventh. Operators who re-estimate only annually are optimising on last year's causality. A deployed uplift model without a holdout control group is not a model, it is a narrative.