Engagement metrics that predict telecom churn risk

Sixty days before the cancellation call, the file already read like a departure: app logins down from eight a month to four, data usage sliding from 12GB toward 4GB, two support contacts in four weeks after nine months of silence. Nobody looked, because none of those numbers appear on a monthly churn report. That gap is the difference between a retention offer that lands and a customer who left mentally in April and administratively in June.

This lesson is about in-life behaviour: which signals move before the decision, how much lead time each one buys you, and how often they lie.

Why lagging churn metrics aren't enough

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 → counts completed decisions, and the money attached to each one is worked through in the retention economics lesson. T-Mobile US has reported postpaid phone churn in the neighbourhood of 0.9% a month in recent years (estimate, from its quarterly investor reporting). On a postpaid phone base in the tens of millions, that steady-looking percentage is still hundreds of thousands of individual exits a month, and the aggregate is stable precisely because it averages away the concentration. Risk is not spread evenly across the base; it clusters in accounts whose behaviour changed weeks ago.

Engagement metrics are the behavioural signals that move first: what the customer does with the app, the loyalty programme, the network and the support channel while still paying you.

The engagement signals that matter most

1. App and portal login frequency

The carrier app is where bills get checked, data gets monitored and plans get compared. A customer who logged in weekly and drops to zero for 30 days has stepped back from the brand relationship, not just the network.

Compare each customer against their own trailing 90-day average, never against a base-wide mean. And control for lifecycle events, because this signal is easy to break: enrol a cohort in autopay and paperless billing and their logins collapse for the best possible reason. If you scored that cohort naively, your model would flag your most locked-in customers as fleeing. Exclude the 60 days after any autopay enrolment or completed device upgrade from baseline comparison.

2. Loyalty-programme redemption

Redemption is the cheapest engagement telemetry a carrier owns: voluntary, timestamped and free to collect. T-Mobile Tuesdays has run weekly giveaways inside the T-Mobile app since 2016; SK Telecom's T Membership discounts at partner merchants have been part of the Korean subscriber relationship for far longer. A customer who claimed three offers in four weeks and has claimed nothing for six has stopped opening the app for the one reason that costs them nothing.

The failure mode is base-wide, not individual. Swap out a headline partner and redemption drops across every cohort at once. Always read redemption relative to the same-week cohort average, so a programme change reads as a programme change rather than 40 million people churning.

3. Data usage slope, not absolute volume

Absolute usage varies too much by plan to threshold. The slope predicts. A sustained month-over-month decline usually means Wi-Fi substitution or, more often now, a second line: dual-SIM handsets and eSIM provisioning let a customer move their real usage to a competitor while your account stays open and billable.

That is the second-order problem. Churn rate never registers this, because nothing cancelled. Usage slope is the only place a silent split shows up, and it typically shows up one to two billing cycles before the port-out.

4. Support contact frequency and cause

Two opposite patterns both carry risk. Silence after a history of engagement suggests disengagement. A spike, two or more contacts in a month after a long quiet period, usually means billing disputes, coverage complaints or plan-comparison questions.

Counting contacts without cause codes produces false positives at industrial scale. One degraded cell site lights up every account in its footprint on the same afternoon, and an untuned model flags a whole postcode as at-risk. Dedupe by geography and known incident window before anything reaches a retention queue. Where you have post-contact survey scores, a detractor rating right after a support interaction is a stronger flag than the contact itself.

Building a composite early-warning score

Individually each signal is noisy. Weighted together they become usable:

churn_risk_score = 
    0.30 * normalize(login_decline_pct) +
    0.35 * normalize(data_usage_decline_pct) +
    0.20 * normalize(loyalty_redemption_decline_pct) +
    0.15 * normalize(support_contact_spike_flag)

# normalize() scales each signal 0-1 relative to
# that customer's own 90-day rolling baseline

Score monthly, flag the top decile, intervene before the customer self-selects into a cancellation request. This is a lightweight version of the propensity-to-churn models carriers build with logistic regression or gradient boosting on the same features. A spreadsheet with four signals and simple thresholds catches most of the value.

The hard part is identity, not maths. App logins sit in one system, usage in the network data warehousedata warehouseA central repository that consolidates data from many source systems into a structured, query-optimized store designed for analytics, reporting, and business intelligence.View full definition →, redemptions in the loyalty platform, contacts in the CRMCRMCustomer Relationship Management: software and strategy to manage and analyse customer interactions throughout their lifecycle.View full definition →. Stitching them to one subscriber is the work, and customer data platformscustomer data platformsSoftware that unifies customer data from every source into one persistent profile that marketing, sales and service teams can act on.View full definition → such as Segment (which sells exactly this plumbing) exist to do it. Decide early whether you score the line or the account: on a family plan the account holder's logins tell you nothing about the teenager whose usage just went to zero. Usage and redemption belong at line level, billing contacts at account level.

Worked example: spotting the 60-day window

A mid-size regional carrier, Account X:

Login decline: baseline 8 to 0, a 100% drop. Usage decline: 12GB to 4GB, 67%. Redemption stopped entirely. The support spike arrived last, at 30 days, which is the point most retention systems wake up.

A rule as blunt as "flag if login decline > 50% AND usage decline > 30%" catches this account at the 60-day mark. The support contact adds nothing you did not already know; by then the customer is comparing prices.

What "good" retention response looks like

Once flagged, the intervention is a marketing decision with a margin attached.

• Proactive discount or data bonus: cuts churn probability among flagged accounts, compresses margin on everyone who was never leaving.
• Plan right-sizing: moving a declining-usage customer down a tier keeps the relationship at lower ARPU instead of losing it, and the value comparison belongs to the LTVLTVLifetime Value: the total revenue (or profit) a customer generates throughout their entire relationship with your business.View full definition → lesson's models.
• Human outreach: for accounts whose modelled value justifies the cost, a call from a retention specialist beats an automated email.

The discipline that most teams skip: hold back a random 10% of flagged accounts and give them nothing. Without that control your reported save rate will look extraordinary, because most flagged customers were going to stay anyway. Incremental saves, not gross saves, are the only number worth reporting upward.

Watch the second-order effect too. If your best offers reliably follow disengagement, some of your base will learn the pattern. Track the re-flag rate: accounts saved twice in twelve months at a discount are not retained, they are repriced. CTIA (US wireless trade association) publishes periodic switching data, though normalising it before you compare is the benchmarking lesson's territory.

Set thresholds by customer type

One threshold across the whole base will misfire.

• Prepaid churn is silent: no call, no port request, just no top-up. Many operators only book it after 60 to 90 days of inactivity, so usage and redemption decay are the only warning you get, and your response window is the recharge cycle, not the contract cycle.
• High-ARPU versus low-ARPU: a 30% usage decline on a 20GB plan costs more in absolute revenue than the same percentage on a budget plan. Prioritise intervention by value, not by risk score alone.
• Market portability speed changes the lead time. In Korea, where SK Telecom competes under fast porting and aggressive handset subsidy competition, disengagement converts to a completed switch quickly. In US postpaid, device financing and trade-in commitments stretch the lag between mental exit and administrative exit, sometimes by months, which buys time and also means a flagged account can sit disengaged for two quarters while still paying.

🎬 [VIDEO: "Customer Churn Prediction Explained" - youtube.com - a walkthrough of how churn prediction models are built from behavioral data, useful for non-technical marketers wanting the intuition without the math]

Key Takeaways

• App logins, loyalty redemption, usage slope and support contacts move 30 to 60 days before cancellation; support contacts move last, so a model built on them alone is already late.
• Measure deviation from each customer's own baseline, and control for lifecycle events: autopay enrolment collapses logins for your most committed customers.
• Usage slope is the only signal that catches a silent dual-SIM or eSIM split, which never appears in churn rate because nothing was cancelled.
• Loyalty redemption drops read as churn only when compared cohort-relative; a partner change moves the whole base at once.
• Hold out 10% of flagged accounts untreated, or your save rate is measuring people who were never leaving, and watch the re-flag rate for customers you have taught to disengage on cue.
• Prepaid gives you no cancellation event and Korean-style fast porting gives you less lead time than US device-financed postpaid, so the response window is a market variable, not a fixed 60 days.