Case Study · WOM Chile · 2022 · Customer Engagement

Churn & Retention
Analytics

A two-year churn intelligence system — from the Dec 2021 CEO executive report through the 2022 weekly dashboards — covering BAM mobile, Fiber home, postpaid, prepaid and business. Cohort analysis, competitive portout intelligence, retention funnel, A/B campaigns, and a churn prediction model that generated zero retentions and taught the team exactly why behavioral signals beat propensity scores.

8.9%4.1% Churn · A/B campaign result
23K+ Retentions in 4 weeks (W44–47)
+2.4pp Share out improvement · Nov 2021
3 Segments · Postpaid · Prepaid · Business

May – Jul 2022 · Weekly · Customer Engagement · BAM Mobile + WOM Hogar (Fiber) · Role: Analytics Engineer

Netezza SQL DWH Datamart Cohort analysis A/B testing Churn modelling CAP traffic correlation Geographic breakdown Highcharts
00 · Prior State · Dec 2021

WOM was outperforming the industry on portout — but the prediction model had zero effectiveness.

The December 2021 executive report to the CEO established the baseline from which the 2022 weekly dashboard system was built. Three findings shaped the analytical priorities for the year ahead: a competitive advantage in portout that was being eroded, a retention funnel that exceeded contact effectiveness targets but revealed a critical model failure, and a growing volume of deactivation intents that the team was only partially converting.

Portout · WOM vs industry vs competitors · Dec 9, 2021
Industry
Portout growth
+14.4%
Context
Market baseline. WOM outperforming.
WOM total
Portout growth
+8.8%
Prepaid
−10.5%
Entel (biggest threat)
Consumer portout to Entel
+38.7%
Top plans at risk
CLP $11K and $13K

WOM's portout was +8.8% vs the industry's +14.4% — already a competitive advantage. But Entel's consumer portout from WOM was growing at +38.7%, disproportionate to its share. This identified Entel as the priority competitor for plan-level retention intelligence, and $11K–$13K as the price band to defend.

Retention initiatives impact · Nov 2021
+2.4pp
Share out improvement from NPG + blindaje retention initiatives in November 2021. Real share out: 20.6% vs 23.0% baseline without interventions. The initiatives were working — the question was how to scale them and fix what wasn't.
00b · Retention Funnel · Dec 2021

The NPG funnel converted 74% of contacts — but the churn prediction model generated zero retentions.

The retention operation ran across four initiative types in W44–47: NPG (negotiation before portout), bajas (deactivation intents from stores and call center), handset renewals, and a churn prediction model. The funnel performance was healthy — until the model results came in. 14,292 clients scored, 2,185 contacted, 0 retentions. Not a targeting problem: the model was surfacing the wrong clients for the retention offer.

NPG Consumer funnel · W44–47 2021
92K
trigger base
NPG
trigger
57K
61.8%
Worked
base
35K
61.7%
Contacted
26K
74.3%
Valid
contact
5.9K
17.5% effectiveness
Retentions
NPG Real

The 74.3% valid contact rate and 17.5% effectiveness both exceeded targets (10% target). But N2+ (second-inquiry) subscribers had 34.8% effectiveness — 2× higher than first-inquiry. This suggested that clients who called twice were more committed to leaving and thus more receptive to an offer. The funnel wasn't the problem. The model was.

Churn prediction model vs manual NPG · same period
Churn prediction model
0
retentions out of 14,292 scored clients. 2,185 contacted. 0 accepted offer. The model identified the correct propensity to leave — but for clients who would not respond to the standard retention offer.
Manual NPG (same weeks)
5,944
retentions from 92K-trigger NPG process. 17.5% effectiveness — exceeding the 10% target. The manual process outperformed the model because it targeted clients in the act of porting, not clients predicted to port.

This is the most important finding in the Dec 2021 report. A churn prediction model with 0% retention rate doesn't mean the model scored wrong — it means the intervention was wrong for that population. Clients predicted to churn in 30–60 days are in a different decision stage than clients who have already called to port. The right intervention for each stage is different. This insight directly shaped how the 2022 weekly dashboards prioritised behavioral signals over modelled propensity scores.

W44–47 retentions · full cross-segment picture
Consumer (total)
20,098
Business
1,997
Prepaid
1,004
Business 5–29 lines A/B · Nov 2021
Control group

−7.57 lines per account (main line). No retention intervention. Business accounts losing lines naturally.

Campaign group

−2.65 lines per account. 434 lines retained in November. 65% reduction in line losses vs control — same pattern as the 2022 BAM data campaign but for the Business segment.

01 · Context

Churn is measured monthly everywhere. It has to be understood weekly to act.

WOM Chile's Customer Engagement team needed more than a monthly churn number. They needed to know which customers were about to leave, why, and which intervention would stop them — segmented by product line, plan, tenure cohort, billing cycle and geography. A single monthly rate is a post-mortem. A weekly intelligence system is a prevention tool.

This case study covers the analytics infrastructure built in 2022 for two product lines: BAM mobile broadband and WOM Hogar fiber. Each had its own churn drivers, its own early-warning signals, and its own retention levers. The dashboards unified them into a single weekly cadence while preserving the product-specific granularity that made the insights actionable.

Before

Monthly churn rate by product. No cohort view, no traffic correlation, no geographic segmentation. Retention decisions made reactively after subscribers had already left.

After

Weekly dashboards with daily churn tracking, cohort heatmaps, CAP-exhaustion-to-churn correlation, geographic breakdown, live A/B campaign performance, and actionable retention proposals with estimated impact in pp.

02 · The Two Product Lines

BAM mobile and Fiber home. Same methodology, different churn drivers.

Monthly churn evolution · BAM consumer
Jul20 – May22 · % monthly churn

The pandemic effect is visible in the cohort data. High churn in 2020 (peak 7.2% Jun-Oct 2020) reflected subscriber activations during lockdown that lapsed as mobility returned. By May 2022 the pattern had stabilised — but the mix had shifted: 54% of churn was now concentrated in subscribers with 0–9 months tenure, a signal that early-life retention was the priority lever.

Churn by type composition · May 2022
voluntary · portout · arrears · fraud
Voluntary
~52%
Portout
~28%
Arrears (mora)
~15%
Fraud / NA
~5%

Portout and voluntary together represent 80%+ of churn — both are preventable. Arrears churn is a different problem (credit risk, not product fit). Separating these in the dashboard meant the retention and collections teams worked off different queues with different logic.

CAP exhaustion rate over time · % subscribers hitting 100% of data cap
Mar21 – Mar22 · avg ~18%

18% of the subscriber base hits their data cap every month. Of those, 70% are on the three highest-cap plans (100 GB, 150 GB, 200 GB) — which are also the three most deviated plans in churn rate. The subscribers most likely to churn voluntarily were also the ones running out of data. That's the retention lever: give them more data before they decide to leave.

Traffic → churn correlation · subscribers with 100% CAP use before churn
~2,400–3,300 churn events/month · CAP distribution at churn moment
CAP exhausters who churn
~18%
of voluntary churn came from subscribers who had used 100% of their data cap in the prior period
Repeat exhausters
35%
of churned CAP exhausters had hit 100% in more than one of their last 5 billing cycles — systematic, not accidental
Zero-traffic churn signal
14%
of subscribers with zero data traffic in one month churned in the next month — disengagement as early warning

Two opposing churn signals, both actionable: subscribers who use ALL their data (frustrated, need more) and subscribers who use NONE (disengaged, need reactivation). A single retention campaign can't address both — the dashboard surfaced both queues separately.

Tenure cohort · churn concentration
months since activation vs churn rate
Vol.
Port.
N3
N6
N12
Trend
[0–3 mo]
17%
10%
25%
15%
8%
↑↑
[4–6 mo]
13%
8%
18%
11%
6%
[7–12 mo]
9%
7%
12%
7%
4%
[25+ mo]
14%
9%
19%
12%
5%

Early-life churn (0–3 months) and long-tenure churn (25+ months) peak simultaneously — two different problems. New subscribers who didn't find product-market fit, and long-tenured subscribers who stopped finding value. The middle cohort (7–12 months) is the most stable. Any retention program should prioritise the extremes, not the average.

Churn evolution · Naked vs Bundle · Oct 2021 – Jul 2022
Naked = Fiber only · Bundle = Fiber + TV

Bundle customers (Fiber + TV) churn at 2× the rate of Naked customers. At first glance this seems counterintuitive — more services should mean more stickiness. The data suggests the opposite: Bundle customers have higher expectations, more points of failure, and are more likely to compare total cost against competitors. Naked churn (1.33–1.69%) was remarkably stable; Bundle churn (1.73–2.77%) showed much more variance.

Voluntary churn Jul22 vs Jun22 · daily tracking
−10.60pp at day 24
Voluntary churn Δ
−0.10pp
Jul vs Jun monthly voluntary churn. Improvement driven by early-life retention focus.
Involuntary churn Δ
+0.14pp
Involuntary churn increased — arrears and non-payment. Different problem, different team.
Budget deviation
+0.60pp
Total churn above budget — surfaced weekly so the team could act before month-close.
Early-life churn · "don't need / don't use" reason
% declaring 'no lo necesita' by tenure band
Month 0–1
~22%
Month 1–2
~18%
Month 2–3
40% of "don't need" → first 3 months
Month 3–6
~12%
6+ months
~8%

40% of "I don't need it" fiber churn happens in the first 3 months. This isn't a pricing problem or a service quality problem — it's an onboarding problem. Subscribers who declare this reason made the wrong purchase decision, often during an aggressive commercial push. The data makes the case for better qualification at point of sale.

Geographic churn leaders · voluntary Jul 2022
% incidencia técnica vs churn voluntario por comuna
San Ramón
98.25% incidencia
Cerro Navia
84.06%
San Bernardo
70.59%
Renca
74.29%
Quilicura
67.44%
Santiago
49.19%

San Ramón (98%) and Cerro Navia (84%) show near-total incidence overlap with voluntary churn. When a network incident's incidence rate approaches 100% of the churning population in a zone, the churn is not "voluntary" in the commercial sense — it's a service quality failure. This allowed the infrastructure team to prioritise remediation zones by their predicted churn impact, not just by ticket volume.

03 · A/B Campaign · Real Result

Give CAP exhausters more data before they leave. Churn dropped from 8.9% to 4.1%.

The traffic analysis led to a direct retention proposal: proactively double the data allowance for subscribers who regularly exhaust their monthly cap. The target population was ~16,500 lines. Instead of rolling it out immediately, a three-variant A/B test was run across different bag sizes to find the most cost-effective dose.

Control group (no intervention)
8.9%
CAP exhausters with no additional data bag. Measured over the same period as campaign groups.
Campaign average (3 variants)
4.1%
Average churn across all three bag-size variants. Halved churn vs control with a data gift — not a price discount.
Campaign variants · bag size vs churn
35 GB bag
4.4%
50 GB bag
4.5%
70 GB bag
3.5%

The 70 GB bag was the most effective at 3.5% churn — but the delta from 35 GB (4.4%) to 70 GB (3.5%) is 0.9pp for double the data cost. The retention decision isn't "which bag works?" — it's "what's the cost per prevented churn at each tier?" The A/B result gave the Finance team the elasticity curve they needed to set policy. Estimated impact: −0.50pp overall monthly churn · ~14,850 subscribers retained out of 16,500 at risk.

04 · What I Owned

Building the intelligence layer that turned churn data into retention decisions.

Decisions I owned

  • Two churn signals, not one. Surfacing both CAP exhausters (overuse) and zero-traffic subscribers (disengagement) in the same dashboard, as separate queues. They require opposite interventions — give more vs re-engage — and combining them in a single campaign would have reduced effectiveness of both.
  • Incidence as a churn cause, not a churn correlate. Joining network incident data to voluntary churn by commune revealed that in zones like San Ramón (98% overlap), "voluntary" churn was actually service-driven. This reclassification changed which team owned the remediation — not retention, but infrastructure.
  • Daily churn tracking vs monthly reporting. Comparing the same day-of-month across consecutive months (e.g. "at day 24 of Jul22 vs Jun22") gave the team an early read on monthly outcome 6 days before close. The −4.8pp reading at day 24 meant corrective action could still be taken within the same month.
  • Cohort heatmap as the primary analytical surface. A cohort × cause matrix (vintage cohort vs churn type vs n3/n6/n12 rate) made the 2020 pandemic distortion visible as a structural artifact — not a benchmark. Without it, 2020-era cohorts would have contaminated trend analysis with a non-representative base period.

Constraints + what I'd change

  • A/B attribution was manual. Campaign group assignment and outcome measurement required manual SQL joins between the campaign list and the churn events table. An automated holdout framework would have made re-running the test for new target populations a configuration change, not a data engineering task.
  • Geographic incidence was lag-matched, not real-time. Network incident data was joined to the prior month's churn, not to concurrent events. Matching incidents to churn events within the same 48-hour window would have increased the signal precision significantly.
  • No churn probability score — only observed churn. The dashboard identified subscribers who had already churned and modelled risk from behavioral signals. A propensity model scoring the full active base before churn events would have shifted the intervention from reactive to truly predictive.
  • Delivery still by PowerPoint. The weekly output was a PPTX sent to stakeholders. A live dashboard (Looker, Tableau, or even a simple web view) would have allowed self-service drill-down and removed the analyst from the critical path for ad hoc questions.
"A churn rate is a fact. A churn rate with a tenure cohort, a traffic signal, a geographic breakdown and a tested intervention is a decision." — Working principle from this project

From revenue assurance and fraud detection to A/B campaign analytics and churn intelligence — consistent discipline: instrument before you need it, and make the output actionable.

Contact Javier