Data Department builds churn prediction, propensity and customer analytics models for automotive, insurance and other consumer-facing businesses, retailers, subscription services, telecom and fintech companies. The models decide which customers to contact, with which offer, and when, so retention and cross-sell budget goes where it actually changes behaviour. For an automotive distribution group, renewal and brand-churn models produced about an 8% revenue gain.
What problems does customer analytics solve?
Marketing and customer teams already know some customers will leave. The hard part is deciding what to do about it. The decisions we work on:
- Which customers to contact this week, before they leave.
- Which offer to give: a discount, a service call, an upgrade, or nothing.
- Which customers would stay anyway, so you do not spend a discount on them.
- Which product to offer next to an existing customer.
- How much a customer is worth over time, and therefore how much to spend keeping them.
- Which customers or accounts to give a higher service level.
A churn score on its own is not enough. The customer most likely to leave may not respond to any offer. The better question is: whose behaviour changes if we act? That is what uplift modelling answers, and it is where much of the revenue gain comes from.
Which model answers which question?
| Model | Question it answers | Typical use |
|---|---|---|
| Churn model | Who is likely to leave, and when? | Weekly retention lists |
| Propensity model | Who is likely to buy a given product? | Cross-sell and upsell campaigns |
| Uplift model | Whose behaviour changes if we contact them? | Allocating offers and discount budget |
| Customer lifetime value | How much is this customer worth over time? | Setting retention spend per segment |
| Recommender | Which product should we show next? | Website, app and email content |
What a project looks like
Discovery (2 weeks). We define churn for your business. For a subscription, it is a cancellation. For a retailer, it is a customer who quietly stops buying, which needs a careful definition. We check customer data and, most importantly, past campaign history: who received which offer and what happened next. You get a go/no-go decision and a pilot design.
Pilot (6–10 weeks). We build the models and run a real campaign with a holdout group that receives no contact. The business metric is fixed on day one, for example revenue retained per contacted customer. Your campaign team runs the contact; we measure.
Production and handover. Scores are refreshed weekly or daily into your CRM (customer relationship management) system or campaign tool. We add monitoring for score drift and campaign results, and train your analysts to read and adjust the models.
What data do churn and propensity models need?
- Customer-level transactions or subscription records, ideally two years or more.
- Usage or activity data: logins, visits, calls, product usage.
- Service interactions: complaints, returns, support tickets.
- Campaign history: offers sent, channels used, responses.
- Optional: customer attributes such as region, tenure or segment.
Campaign history matters more than most companies expect. Without it, a model can say who is at risk but not which action works.
Results from real projects
For an automotive distribution group, two linked models estimated which customers were likely to replace their vehicle soon and, of those, which were likely to switch brands. The scores decided which customers to call first and which price offer to make. The result was about an 8% revenue gain (vehicle renewal case study). For a mobile game, a day-1 retention model reached an AUC (area under the curve) of 0.82, enough to flag new players likely to leave after their first day.
Customer models work well alongside dynamic pricing, because offer depth and price sensitivity are closely linked. See how they fit together for retail and e-commerce.
When is churn prediction NOT the right tool?
- You cannot act on the score. With no contact channel, no offer budget or no team to run campaigns, a churn list changes nothing.
- The cause is obvious and shared. If everyone leaves because of one product problem, fix the product first.
- The customer base is very small. With a few hundred customers, account managers usually know who is at risk.
- There is no legal basis for contact. Without marketing consent, a model cannot help.
Frequently asked questions
How much data do we need for churn prediction?
Usually a few thousand customers with at least a year of history, including some who left. Campaign history makes the models much more useful.
How long until we see results?
Discovery takes 2 weeks and the pilot 6–10 weeks, including a live campaign with a holdout group. You see measured results before deciding on production.
Can the models run inside our CRM?
Yes. We deliver scores into your CRM, campaign tool or data warehouse through database tables, file exports or APIs (application programming interfaces).
What drives the cost of a customer analytics project?
The number of data sources to join, how clean customer identifiers are across systems, how many models you need (churn only, or churn plus propensity and uplift), and how often scores must refresh.
How do you handle KVKK and GDPR?
We work with pseudonymised data on your infrastructure, use only the fields the model needs, and follow your consent records for contact. This keeps the work within KVKK (Türkiye's Personal Data Protection Law) and GDPR (General Data Protection Regulation).