Customer Churn Prediction for Telecom

Machine learning model trained on your subscriber data that scores each account's probability of churning within a defined prediction window, typically 30 or 60 days. Feature engineering draws from CDR (Call Detail Record) data to extract usage pattern trends: declining data consumption volume over rolling 4-week windows, call frequency reduction, SMS volume changes, and roaming usage shifts. Network quality features per subscriber are derived from RSRP (Reference Signal Received Power) and SINR (Signal-to-Interference-plus-Noise Ratio) measurements to capture the individual signal quality experience that aggregate network metrics miss. Bill shock events -- months where the invoice exceeded the subscriber's average by a defined threshold -- are included as point-in-time features. Support call volume, complaint category, and resolution outcome are extracted from CRM interaction logs. Device age, handset upgrade recency, and contract tenure round out the feature set.

The model architecture uses gradient-boosted trees -- XGBoost or LightGBM depending on dataset characteristics -- with Platt scaling calibration applied to the raw output scores so the predicted probability is reliably calibrated and comparable across subscriber segments. The model is trained on your historical churn labels (churned versus retained over a matched observation window) and validated on a held-out test set. Propensity scores are updated on a configured schedule (weekly batch or daily incremental) for your full active subscriber base.

For each subscriber's risk score, the model surfaces the top contributing factors that drove their score above the at-risk threshold. SHAP (SHapley Additive exPlanations) values are computed per subscriber to identify which features contributed most positively or negatively to the churn score, with the top three to five factors displayed in the retention agent's view of the subscriber record.

A subscriber scored as high churn risk because their data usage dropped 40% over four weeks, they had two service quality complaints in the last month, and their contract expires in 45 days warrants a different retention message -- likely a service quality follow-up combined with a renewal offer -- than a subscriber whose risk is driven by price comparison browsing signals and a recent competitor switch offer, who would respond better to a price-matching or loyalty discount. Contributing factor labels are translated into plain language in the agent UI so the retention team reads "data usage dropped sharply in the last 30 days" rather than a feature code. This makes the risk score actionable and keeps the outreach relevant -- a retention call that addresses the wrong problem is worse than no call, because it trains the subscriber that the operator doesn't understand their experience.

Churn risk prioritised by subscriber lifetime value so retention investment focuses where it delivers the highest financial return. LTV is calculated from current ARPU, remaining contract tenure, predicted post-contract renewal probability, and historical churn rate for subscribers in the same plan and tenure cohort. A high-value subscriber on a premium plan with high ARPU, long tenure, and no prior churn history scores differently in the retention queue than a low-value subscriber approaching contract end on a basic plan, even when both show the same raw churn propensity score.

The combined risk-LTV matrix segments the at-risk subscriber population into treatment tiers. Tier one -- high churn risk, high LTV -- gets immediate agent outreach with a premium retention offer. Tier two -- moderate churn risk, high LTV -- gets a proactive communication with a softer retention nudge. Tier three -- high churn risk, low LTV -- goes to an automated SMS or app-push retention journey without agent involvement to preserve capacity for tier one and two. Uplift modelling using a Bayesian bootstrap T-learner can further refine the tier one population by identifying treatment-responsive churners -- subscribers who will respond to a retention intervention -- and filtering out those who would churn regardless of what offer is made. This improves the cost-effectiveness of the retention programme by avoiding wasted offer spend on subscribers who are already lost.

Automated retention interventions triggered by churn risk scores without manual agent involvement for lower-risk and lower-value segments. Intervention channels are selected based on the subscriber's tier: SMS for mobile-primary subscribers, in-app push notification for subscribers with high app engagement, email for subscribers with lower mobile engagement, and agent queue assignment for the highest-value at-risk accounts. Offer personalisation is driven by the contributing factor analysis -- a subscriber whose risk is driven by network quality complaints receives a service credit or speed upgrade offer rather than a price reduction that doesn't address the actual dissatisfaction.

Offer acceptance, rejection, and non-response are tracked per subscriber and per campaign. Acceptance data feeds back into the model as a training signal for future intervention targeting. For high-value at-risk subscribers, the agent alert includes the subscriber's risk score, top contributing factors, tenure, ARPU, and a suggested retention approach pre-populated from the playbook you define during setup. Integration is built with your CRM (Salesforce, HubSpot), contact centre platform, or marketing automation system so at-risk subscriber lists and intervention outcomes flow between the prediction system and the tools your team already uses. The intervention layer is what makes the model commercially useful -- a score without an intervention workflow produces reports, not revenue.

Retention programme dashboards built for commercial leadership and the retention operations team, with different views for each audience. The commercial view shows current at-risk subscriber count by segment and value tier, predicted churn volume for the next 30 days, estimated revenue at risk, and retention programme ROI calculated from offer cost against retained revenue. The operations view shows at-risk subscriber lists by intervention status, agent workqueue depth, and intervention campaign performance metrics.

Model performance metrics are tracked and displayed: AUC-ROC (target 0.75 or higher), precision at the 10% recall threshold, and lift versus random contact -- all updated each time the model runs against fresh data. Churn rate trend versus prediction gives the operations team visibility into whether the model's predicted churn volume is tracking against actual churn, which is an early warning of model drift before it affects programme performance. Retention intervention performance metrics show how many at-risk subscribers were contacted, which offers were made, acceptance rate by offer type, and the 60 or 90-day retained rate for contacted versus non-contacted at-risk subscribers. The dashboard is the tool that makes the programme's performance reviewable and improvable over time.

Churn prediction models degrade over time as subscriber behaviour patterns change -- new competitor offers enter the market, network quality improvements reduce complaint volume, handset upgrade cycles shift, or economic conditions change spending and contract commitment patterns. Without active monitoring, the model continues producing scores that are increasingly disconnected from current behaviour, and the retention programme loses effectiveness without any obvious signal that the underlying model is the cause.

We design the model with a monitoring layer that tracks PSI (Population Stability Index) monthly on key input features, with a threshold of 0.2 as the alert level for significant distribution shift. AUC-ROC and recall on a held-out validation set are tracked at each scoring run. When PSI breaches the threshold on two or more key features, or when model performance drops below a defined minimum, a retraining pipeline triggers automatically. The pipeline ingests fresh training data from the most recent observation window, retrains the model, validates performance against the hold-out set, runs a shadow comparison against the current production model, and promotes the new model to production if it outperforms the incumbent. The full monitoring and retraining design is built during the initial project so the system doesn't require manual intervention each time the world changes.