Technology
Designed to be right, not to look impressive.
Churn models can easily look better than they are. Churnico's technology approach focuses on the one thing an enterprise decision needs: a prediction that works in the real world.
Three layers, one source of truth.
- L3
Decision and language layer
Scores become an action list and dashboard. The natural-language assistant never calculates; it only explains validated results.
- Action matching
- Dashboard
- Language assistant
- L2
Modelling layer
Gradient-boosted ensembles work alongside a business rule engine. Probabilities are calibrated; every score is produced with its drivers.
- Ensemble models
- Rule engine
- Calibration
- Explainability
- L1
Data and feature layer
Sources are joined on a customer-time axis. Every feature is built only from information available at prediction time.
- Point-in-time features
- Data quality gates
- Anonymous identifiers
The discipline that protects model quality.
- 01
Out-of-time validation
The model is tested on a real period after the one it learned from. Cross-sectional and real deployment performance are reported separately.
- 02
Leakage audits
Fields that already 'know' the outcome, such as records created after cancellation, are found with systematic tests and removed.
- 03
Calibrated probabilities
A 0.8 score means customers with this profile really do leave at a high rate, so thresholds can be read with confidence.
- 04
Explicit limits
Questions the data can't answer are marked 'no data'. Missing metrics are never estimated or written as zero.
- 05
Rules and models together
When simple, robust business rules outperform a complex model, seeing that and using it is part of the modelling job.
- 06
A language model that doesn't calculate
The assistant never produces numbers; it selects pre-computed values. Every number in an answer is automatically checked against its source.
Data responsibility
Designed to touch no more customer data than needed.
Anonymous identifiers
The model works with customer IDs that carry no personal information.
Data minimisation
Only fields that contribute to the churn question are requested.
Regulatory alignment
Data protection requirements and processing terms are agreed at the start.
Deployment options
Data flow and hosting are set according to your security policy.
See your churn risk in your own data.
Let's talk through your data in a short call and assess together which signals Churnico can read in your customer base.