Predictive Segmentation

Predictive Audiences · ML Segmentation

Definition

Predictive segmentation groups users by machine learning predictions of their future behaviour — churn risk, purchase probability — allowing pre-emptive action instead of a reaction after the fact.

What predictive segmentation is

Classical segmentation groups users by what already happened: demographics, purchase history,
on-site behaviour. Predictive segmentation uses machine learning to answer a different question:
what will this person do next?

In practice that means three kinds of prediction:

Prediction	Question	Use case
Churn prediction	Will this shopper leave in the next 30–60 days?	A win-back campaign before they go
Purchase probability	Will they buy in the next 7–14 days?	Hot leads for a promotion
LTV prediction	What is the expected annual revenue from this customer?	Prioritising the VIP segment

Churn prediction: how the model works

A churn model studies the behavioural patterns of shoppers who left in the past and looks for
similar patterns among active ones:

Falling visit frequency (four times a week before, once a month now)
Shorter, shallower sessions
Disengagement from recommendations
Smaller orders, or a shift to a cheaper segment

The output is a score between 0 and 1 for each shopper. Those above a threshold such as 0.7 enter
the risk segment.

Important: a predictive model is probabilistic, not deterministic. Someone in the high risk
segment is not a person who will definitely leave, but one where pre-emptive action is
statistically justified.

Putting predictive segments to work

The churn segment: run a personalized win-back campaign focused on the categories the shopper
has affinity for, not on generic promotions. The goal is to restore engagement before the interest
disappears entirely.

High purchase probability: focus personalization resource on the shoppers who are close.
Show them precise recommendations rather than bonus offers — they were already near the decision.

High predicted LTV: identify future VIP customers early and invest disproportionately in keeping
them.

Data requirements

Predictive models need enough history:

At least 6–12 months of transaction history
Behavioural events (sessions, views, clicks)
A base of 10,000–20,000 users with transactions for a reliable model
Regular retraining — behavioural patterns shift with seasonality and market trends

Frequently asked questions

How does predictive segmentation differ from RFM?

RFM is retrospective: it groups by what already happened — when someone last bought, how often, how much. Predictive segmentation looks forward: a model computes the probability of the next purchase or of churn from behavioural patterns. A shopper can have excellent recency and still carry a high churn probability if the trajectory of their behaviour points downward.

What data does predictive segmentation need?

At minimum: three to six months of transaction history and behavioural events on the site (views, clicks, sessions). More history means a better model. The first reliable predictive segments appear on a base of 10,000–20,000 users with transactions.

How should a high churn risk segment be used?

Run a win-back campaign before the churn happens: a personalized offer, content from the categories the shopper prefers, a discount in a high-affinity category. This segment has to be worked two to four weeks before the expected silence, not after it.

What is purchase probability and when will it ship in Gravity Field?

Purchase probability predicts the chance that a specific shopper buys within the next N days. It allows an audience to be segmented by readiness so budget goes to those close to a purchase rather than to cold traffic. In Gravity Field churn prediction is available; purchase probability is on the roadmap.

Gravity Field