How did Telefónica build a churn model that actually moved retention numbers?

Telefónica's Spanish operation faced a problem that every major integrated operator knows well: aggregate churn rates in the low single digits mask enormous variation underneath. A prepaid customer who stops topping up looks identical in the billing system to a high-value postpaid subscriber who has quietly ported their secondary line to a rival. Both register as churned. Neither should have been managed the same way in the weeks before they left.

By 2022, Telefónica's data science group in Madrid had concluded that billing data alone, which was the primary input to their earlier propensity models, was systematically too slow. A subscriber who has already decided to leave stops paying before they stop using the network, so the churn signal arrives after the intervention window has closed. The team reoriented the model architecture around network-layer signals that move faster than any billing event.

How Telefónica rebuilt its churn signal architecture around CDRs and network telemetry

The core shift was moving from event-driven billing records to continuous network telemetry. Call detail records capture call completion rates, handoff failures, and roaming transitions in near real time. When a subscriber's device starts attaching to a rival's roaming partner more frequently, or when voice call setup failures spike on a particular cell sector serving that subscriber's home location area, those patterns appear in the RAN telemetry hours before they appear anywhere in a CRMCRMCustomer Relationship Management: software and strategy to manage and analyse customer interactions throughout their lifecycle.View full definition → record. For a deeper look at the raw material involved,the full taxonomy of CDR fields and network telemetry streams shows why billing data covers only a fraction of what the network actually knows about a subscriber.

Telefónica structured the feature set around four signal categories: usage trajectory (month-on-month decline in data consumption, shift from voice to OTT substitutes), network experience (packet loss rates, video stall ratios, 4G-to-5G upgrade eligibility versus actual upgrade rate), commercial interactions (contact center calls about billing disputes, failed self-service transactions in the app), and competitive context (postcode-level NPSNPSNet Promoter Score (NPS) measures customer loyalty by asking how likely customers are to recommend a brand, then subtracting detractors from promoters.View full definition → movements correlated with a rival's new site activations). The competitive context layer required pulling in external data, specifically Opensignal benchmarks and CNMC regulatory filings in Spain, to proxy where a rival was improving coverage.

The model itself was a gradient boosting ensemble, trained on 18 months of labeled churn events and retrained monthly. The retraining cadence mattered: telecom competitive dynamics shift fast enough that a model trained before a major competitor's price promotion can misrank the population it is supposed to protect. Telefónica's MLOpsMLOpsMachine Learning Operations: combining ML and DevOps practices to industrialise, deploy, monitor, and retrain models reliably in production.View full definition → team built automated drift detection to flag when the predicted churn score distribution diverged more than two standard deviations from baseline, triggering an out-of-cycle retrain.Maintaining that kind of production discipline across a model that scores millions of subscribers weekly is a different engineering problem from building the model in the first place.

One constraint that shaped architecture decisions throughout: Spanish ePrivacy rules and GDPRGDPREU regulation governing how organizations collect, store and use personal data, with fines tied to global revenue for breaches.View full definition → Article 22 requirements. Automated decisions with legal or similarly significant effects require human review or explicit consent to opt in. Telefónica's legal and data teams determined that a personalized retention offer generated by a churn model did not cross the Article 22 threshold as long as a human agent reviewed the recommendation before contacting the customer. That determination was documented in the model's governance register and audited annually, not assumed once and forgotten.

Did Telefónica's churn model actually reduce voluntary churn?

The honest answer is: probably yes, but the attributionattributionA framework for assigning credit to the touchpoints that contributed to a conversion, so you can measure which channels and interactions actually drive results.View full definition → is complicated. Telefónica's annual reports from 2023 and 2024 reported improved postpaid churn figures in Spain, with postpaid churn declining toward 0.8-0.9% per month in the core Spanish market, below the Western European operator average of roughly 1.2-1.4% at the time. Telefónica has publicly attributed part of this improvement to improved targeting in retention campaigns. They have not published a controlled experiment result isolating the model's contribution, so a precise causal figure is not available.

What their data teams have described in public conference presentations is a shift in retention spend efficiency: fewer blanket discount offers sent to subscribers who would have stayed regardless, and higher-value interventions concentrated on the population the model scored as genuinely at risk. The cost-per-save metric improved more than the raw churn ratechurn rateChurn rate is the percentage of customers or revenue lost over a period. It measures how fast a business loses its existing customer base.View full definition →, which matters in a sector where retention discounts directly compress ARPU.

What transfers to other telco CDOs, and where context diverges

The signal architecture Telefónica used transfers directly to any operator with access to RAN telemetry, which means any operator running their own network. A pure MVNO reselling capacity on a host network does not see packet loss rates or handoff failures for their subscribers; they see only what the host passes through billing feeds, which is close to what Telefónica had before their rebuild. That is a structural constraint on model quality, not a data engineering problem.

The competitive context layer is harder to replicate cheaply. Opensignal benchmarks are a commercial product. CNMC regulatory data is public but requires significant processing to reachreachThe number of unique people exposed to your message in a given period. Unlike impressions, reach counts each person once, no matter how often they see it.View full definition → postcode-level usability. Smaller operators often skip this layer and then wonder why their model underperforms in postcodes where a rival has been quietly upgrading infrastructure.

Three decisions any CDO in this position needs to make explicitly before building:

• Define the intervention window. If your retention team can call a customer within 48 hours of a score crossing a threshold, you need signals that predict churn 2-4 weeks out. If your contact center operates monthly batch campaigns, you need a longer prediction horizon and accept more noise.
• Decide what counts as churn in the training labels. Port-out is clean. Voluntary disconnection is clean. A subscriber who downgrades to a free SIM-only plan and reduces usage by 80% may be economically equivalent to churn but will not appear in your labeled dataset unless you build a "silent churn" category explicitly.
• Build the governance structure before model deployment, not after. The ePrivacy and GDPR constraints Telefónica navigated are not unique to Spain; any EU operator faces the same framework, and operators in markets with GDPR-equivalent legislation (Brazil's LGPD, UK GDPR post-2020) face analogous rules.

Churn prediction in telecom is not a modeling problem that gets solved once. Competitive dynamics, network topology changes, and regulatory interventions (number portability rule changes, wholesale access decisions) all alter the relationship between behavioral signals and actual churn events. The operators who maintain a retention edge are the ones that treat model refresh as an operational process, not a project milestone.

The full course on this sector:Data in Telecom.