Case study
Retail predictive analytics case study: lapse risk and customer lifetime value for a specialty retailer
ResultLoyalty-member lapse scores, value bands and explanations were delivered to the campaign tool, with offers controlled by the CRM team.
- Client
- A US specialty retailer (anonymised)
- Industry
- Retail (home and outdoor goods)
- Service
- Predictive analytics: a lapse model and a lifetime value model
- Type
- Two scores on one set of member features, written into the campaign tool
- Status
- Delivered
We built predictive analytics for a specialty retailer's loyalty programme. The project combined lapse-risk scoring and customer lifetime value, then placed the scores and their explanations in the tool where the CRM team builds campaigns. Campaign measurement included a held-out audience so the team could distinguish model ranking from the effect of an offer.
The starting point
The retailer used time since the last purchase to trigger win-back campaigns and previous spending to select higher-value offers. Those simple rules treated members with different shopping patterns alike: someone at risk of lapsing could look the same as someone whose usual purchase cycle was longer. Revenue ranking also missed the effect of returns on customer value.
What we delivered
We defined lapse and value with the CRM team and built member features using only information available at the prediction date. A calibrated gradient-boosted model scores lapse risk, while the value model estimates purchase frequency and spend with margin after returns as the business measure. Dated scores, bands and reasons feed the warehouse and campaign tool.
Rollout and handover
We compared the models with the existing rules on the same historical periods and checked feature leakage and calibration. Campaign integration included marketing opt-outs and a held-out group excluded from offers. The operating process covers scoring, stability checks and retraining, with the original rule available as a fallback.
Where people remain in control
The models score members; the CRM team selects offers, sets discounts and approves contact. Marketing opt-outs and the held-out audience are excluded from campaign targeting.
Project outcome
The CRM team received member-level scores and explanations alongside its campaign audiences. It retains control of offer selection, discounts and contact. The measurement framework separates how well the models rank members from whether a campaign produces additional purchases and margin.
How performance is evaluated
| Measure | How it is measured | Comparison point |
|---|---|---|
| Ranking (AUC) | The lapse model trained on earlier months and tested on the newest month with known outcomes | The previous win-back rule, on the same month |
| Calibration | Members grouped by score, with each group's predicted lapse rate set against the rate that followed | The previous rule gave no probability |
| Lift at the mailing budget | The lapse rate among the members the monthly budget can reach, against the rate across all members | The rule's list of the same size |
| Stability | The ranking test repeated on several past months, one after another | Comparison across evaluation periods |
| Value ranking | Whether members in the highest predicted-value band bring the most margin in the months that follow | Ranking by last year's spend, as the rule does |
| The action test | Purchases and margin over the agreed observation window, offered members against the held-out group | The held-out group |
Technologies and tools
- The retailer's cloud data warehouse
- A lapse model: gradient-boosted trees, calibrated
- A lifetime value model of purchase frequency and spend
- Scores, bands and reasons in the campaign tool
- Monitoring tables for the five gates
More case studies
All case studies- Case studyWeekly supply-planning forecasts for a packaged food makerWeekly product and distribution-centre forecasts were connected to supply planning, with comparable accuracy reports and planner overrides.
- Case studyField stations that identify birds by sound, with set-up cut to under 10 minutesAverage station set-up: from nearly 25 to 30 minutes to under 10
Bring us the decision your team wants better evidence for
Tell us what you want to predict, how your team chooses an audience today and where a score would need to appear. An engineer replies within two business days with whether your history could improve on the current rule and how we would test it.