Most customer lists rank people by what they already spend. That identifies value, but it does not answer the decision a promotion needs to make: whose behavior is likely to change because of this action, and what will that change cost?
The interactive model on this page is an authored comparison of customer states and interventions. Its 0 to 100 indices are Kelvin-selected assumptions. They are not observed lift, causal estimates, or production results.
DECISION FRAME
Value and incremental response are different
High purchase likelihood identifies a valuable customer. Incremental response asks whether an intervention changes what that customer does. Each customer sits at their baseline if the shop holds, and rises by the modeled response to the action you choose. Pick a customer to send them to the model.
Little discount wasteMost of the discount wasted on a purchase that would have happened anyway
Lapsed regular · Targeted offerRepeat likelihood 22 if the shop holds, 74 under this action: a modeled response of +52 with discount waste 24. Highest recovery per dollar of the four, because the habit is still recent.
The question the system answers
A steady regular can be valuable and still be a poor discount target. If that customer already buys every three weeks at list price, a discount may change the price without changing the purchase.
The system asks three questions in order. Is the customer drifting from their own pattern? When would the next purchase normally happen? Is money required to change the outcome, or would a timely message do the same job?
That sequence protects the distinction between baseline purchase likelihood and incremental response.
Why spend ranking picks the wrong customer
Spend ranking puts steady regulars and promotion-trained customers near the top of the same list. They need different treatment.
The steady regular has high baseline purchase likelihood and little reason to receive an offer. The promotion-trained customer may respond to a discount, but each response can deepen the expectation that another discount is coming. The lapsed regular is often the more useful test population because a real habit existed and recently broke.
The model uses 20 percent and 60 percent discounted-order shares as deliberately selected operating thresholds. Below 20 percent, it treats behavior as largely independent of promotions. Above 60 percent, it treats behavior as promotion-led. Those boundaries are heuristics to test, not thresholds learned from customer data.
Prior art and the claim boundary
This decision frame is related to uplift modeling and heterogeneous treatment effects. Those methods estimate how an outcome changes under a treatment, conditional on customer attributes, rather than only predicting who is likely to purchase. Microsoft Research's EconML and the CausalML methodology describe that causal question.
My contribution here is the operating model: how recency against a personal baseline, interval stability, promotion dependence, and margin headroom can route a customer into hold, reminder, targeted offer, or deeper offer decisions.
Purchase history alone cannot establish incrementality. The scores on this page express a decision hypothesis. A randomized treatment and holdout design is required before any response difference can be called lift.
How to read the model
Choose a customer situation and an intervention. Every pairing has four authored indices and a stated failure condition. Move the steady regular across all four actions. Repeat likelihood changes only slightly while discount waste rises sharply. Then move the lapsed regular across the same actions and compare the trade.
The model is useful when it makes the cost of a choice visible. It is wrong if a high score is treated as a measured probability.
INTERACTIVE MODEL
Compare customer state and intervention
Change the customer situation and intervention to inspect an authored decision hypothesis.
Illustrative indices. These are Kelvin-selected scenario assumptions, not measured response rates, causal lift, or probabilities.
THE BRIEF'S EXPERIMENT
Move the steady regular across all four actions. Repeat likelihood changes only slightly while discount waste rises sharply. Then move the lapsed regular across the same actions and compare the trade.
Steady regularPredictable interval, buys at list price, rarely touches a promotion.
Targeted offerA modest discount released at the decision moment rather than on a calendar.
Showing Targeted offer for Steady regular. Repeat likelihood 86 if the shop holds, 90 under this action. Repeat likelihood 90, Margin retained 79, Discount waste 88, Relationship health 74.
ILLUSTRATIVE INDICES / 0 TO 100
Against holding
What does this action change for this customer, compared with doing nothing?
+4 modeled response in repeat likelihood, bought with discount waste of 88 and margin retained of 79.
This pairing
Higher suggests the customer makes another purchase under these assumptions.
Higher means more of the sale's contribution survives the intervention.
Higher means more money went to a customer who would have purchased anyway.
Higher suggests the contact reads as useful rather than as pressure.
Response map
All twenty pairings by modeled response over holding and discount waste. The dashed line joins this customer's four options; click one to choose it.
What this buys you
Marginally faster repurchase and a small basket increase.
What it costs you
Repeat likelihood moves about four points while most of the discount lands on a sale already won.
All four actions for the steady regular
| Action | Response | Repeat likelihood | Margin retained | Discount waste | Relationship health |
|---|---|---|---|---|---|
| Control | Control | 86 | 100 | 0 | 82 |
| +2 | 88 | 100 | 4 | 76 | |
| +4 | 90 | 79 | 88 | 74 | |
| +6 | 92 | 44 | 97 | 66 |
BEFORE ANY OF THIS IS CALLED LIFT
- PRIOR ARTThis decision frame is related to uplift modeling and heterogeneous treatment effects, which estimate how an outcome changes under a treatment rather than only predicting who is likely to purchase. EconML CausalML methodology
- HEURISTICThe model uses 20 percent and 60 percent discounted-order shares as deliberately selected operating thresholds. Those boundaries are heuristics to test, not thresholds learned from customer data.
- REQUIREDPurchase history alone cannot establish incrementality. A randomized treatment and holdout design is required before any response difference can be called lift.