CUSTOMER RETENTION
Churn prediction: find the customers about to leave
Churn prediction with TIMi scores each customer on the risk of leaving and shows the variables behind that score, so retention teams act on the real cause instead of a hunch.
Prepared in Anatella
History, contracts and interactions from 80+ sources, turned into model-ready variables.
Modeled in Modeler
AutoML builds and compares models on your full customer base, with no sampling.
Explained per customer
Every score comes with the variables that weigh most for that customer.
HOW IT WORKS
How churn prediction works.
Three steps, in one engine, on a standard server with no cluster.
1. Prepare the data
Anatella gathers transaction history, contract data and customer interactions, cleans them and turns them into variables the model can use.
2. Build the model
Modeler generates candidate models with AutoML and keeps the one with the best balance of accuracy and interpretability. The retained model typically relies on 15 to 25 variables, picked among thousands of candidates.
3. Score and explain
Each customer gets a churn risk score and the variables that weigh most in it for that customer, not a global average.
GETTING IT RIGHT
Three decisions that make or break a churn model.
Most of the work is in how the target is defined, not in the algorithm.
Fix the observation date and the prediction window
The model learns from customer profiles as they were at a past date, and from who left in the period that followed. In telecom, a prediction window of 3 months is common. If your systems do not keep history, snapshots of the customer base can be created at regular dates.
Separate commercial and financial churners
Commercial churners leave by choice, when their contract ends. Financial churners leave early because they can no longer pay. Offering a retention incentive to financial churners can lower the retention rate, so the two groups are usually modeled apart.
Leave out the consequences of churn
A column such as the termination date only exists once the customer has left. If it stays in the dataset, the model learns the consequence and not the cause. Finding these columns takes a large share of the modeling time, and Modeler lets you exclude a column in one click.
EXPLAINABILITY
A score says who. The drivers say why.
A risk score alone tells you a customer may leave. A white-box model also shows, variable by variable, what pushes that score up for this customer, such as a late invoice, a drop in usage or a recent contact with support.
For retention teams
The retention action targets the actual cause of the risk instead of a generic assumption.
For transparency requirements
A model that justifies each result variable by variable gives you the documentation to answer transparency expectations such as GDPR and the AI Act. The compliance assessment itself stays with your legal team.
For model owners
Modeler’s built-in Drift Report shows when a model loses accuracy, so you update it when needed instead of on a fixed calendar.
DATA
What data a churn model uses.
Anatella connects to more than 80 sources, so this data can be gathered without prior custom development.
Transaction history
Frequency, amount, regularity and recency of purchases or payments.
Contract data
Tenure, offer type, binding period and past contract changes.
Customer interactions
Support requests, complaints and the channels used to contact you.
Usage signals
How often the product or service is used, and how that evolves.
PROOF
Seen in production.
Results from our published customer stories.
Attrition cut from 10% to 8% a year
Contacting the riskiest business clients identified by the model saves 7,000 more professional accounts every year.
Churn detection accuracy raised from 45% to 57%
MTN analyzed 6 months of data for 21.4 million subscribers in under 5 days, using Social Network Analysis.
1,013 variables flag churners across 37 categories
Delhaize detects disengagement early enough to act, and pairs detection with relevant offers.
FAQ
Common questions.
How long does it take to build a first churn model?
Most of the effort goes into preparing the data, not into modeling. The timeline depends on how accessible and clean your sources are, which we review with you before starting.
What is the difference between commercial and financial churn?
Commercial churners leave by choice, when no contract binds them anymore. Financial churners leave early, even if they pay a penalty, because they can no longer afford the service. The first group responds to retention efforts, the second usually does not, so they are modeled separately.
Do we need a dedicated data scientist?
Modeler generates and selects models automatically, and its wizard interface is designed for users without a data mining background. An analyst is still useful to define the target and to read the results with the business teams, but not to build the model by hand.
Does the model need to be rebuilt when customer behavior changes?
Churn models lose accuracy as behavior shifts. Modeler includes a Drift Report that shows when a model has degraded, so you update it when needed instead of on an arbitrary schedule.
Can relationships between customers improve the prediction?
Yes, when you have a relationship graph, such as calls between subscribers. Anatella includes Social Network Analysis operators that turn it into variables, for example the number of churners around a customer. MTN used this approach.
RELATED USE CASES
Keep exploring.
Customer Lifetime Value Prediction
Estimate what each customer is worth over time, and what drives that value.
Next Best Offer
Rank the offers each customer is most likely to accept, with the reasons why.
Cross-Sell Prediction
Find the customers most likely to buy a complementary product, and why.
SEE FOR YOURSELF
Test it on your own data.
Run TIMi on your own data with a free trial, or talk to us about your case.
