FTTH growth and retention
Broadband Customer Churn Prediction: Predict Only When You Can Act Profitably
Knowing which subscribers may leave is useful. Knowing which intervention can change their decision, at a sustainable cost, is what turns broadband customer churn prediction into a commercial advantage.
A ranked list of risky accounts can create expensive busywork. Teams call households that would have stayed, discount plans without fixing service problems and report accepted offers as “saves.” The score looks productive while the campaign’s incremental value remains unclear.
For FTTH leaders, the better starting point is a decision: which customer problem can we resolve, who owns the response and what evidence would justify the spend?
What is broadband customer churn prediction?
Broadband customer churn prediction estimates the likelihood that a subscriber will leave within a defined period using information available before that period. It can help identify early-warning signals, but a churn score does not show whether a campaign will keep the customer. Profitable targeting also requires customer value, intervention costs and evidence of additional retention caused by the action.
Use prediction to support an owned, testable intervention. Judge the result through incremental contribution and lasting retention, alongside offer acceptance and conversion.
Current research: good prediction is not proof of profitable retention
A Scientific Reports study published on 12 December 2025 reported 95.13% prediction accuracy and an AUC of 0.89 on a public telecom dataset containing 2,668 customer records. AUC describes how well the model separates churners from non-churners across score thresholds. The paper also reports churn recall of 0.68, meaning that approximately 32% of actual churners were missed at its evaluated threshold.
Those are model results, not measured campaign ROI. The dataset’s call-related predictors also make the results unsuitable as a direct benchmark for an Indonesian FTTH customer base. Leadership still needs evidence that an intervention changes outcomes.
A complementary telecom study introduced an Orange Belgium dataset for uplift modeling, which estimates how an intervention changes behavior. Its distinction is commercially important: “likely to leave” and “more likely to stay because we act” are different questions.
Both sources were checked on 3 October 2026 and are linked below. The first is a peer-reviewed 2025 study; the second is a 2023 research benchmark. Neither establishes a universal broadband accuracy target or an expected retention gain for your ISP.
Start with early-warning signals the CMO can act on
Begin with billing, support, service-quality and lifecycle information already available to your team. The signals below are hypotheses to test, not universal predictors. Combine them with customer context before assigning a reason for churn.
| Signal | What to verify | Relevant response | Owner |
|---|---|---|---|
| Repeated complaints | Is the same fault recurring or was the previous case unresolved? | One recovery owner, a verified fix and a follow-up on the original problem | Support lead |
| Degraded experience | Access-line fault, congestion, home Wi-Fi or device limitation? | Diagnostics and appropriate service recovery before a sales offer | Operations and support |
| Failed payment | Transaction error, affordability pressure or deliberate nonpayment? | Clear payment help and suitable account options | Billing lead |
| Plan mismatch | Does the household need different capacity, coverage or pricing? | A justified upgrade, downgrade or coverage assessment | Sales and product |
| Declining engagement | Travel, changed usage, alternate connectivity or reduced need? | A useful check-in, without assuming low app usage means dissatisfaction | Lifecycle marketing |
Personalization means addressing the verified need. A family with recurring Wi-Fi dead zones needs a diagnostic route; a subscriber with an incorrect bill needs the bill corrected. Neither should receive the same generic discount email.
For the broader service and commercial context, read our perspective on FTTH retention across price, reliability and Wi-Fi experience. Prediction helps prioritize investigation; it does not replace fixing the underlying problem.
Rank by the value of an effective action, not risk alone
The Chief Sales Officer needs three separate estimates: churn risk, offer acceptance and incremental retention. Acceptance helps plan contact capacity and incentive cost. It is not evidence that the customer would have left without the offer.
Where campaign evidence is available, prioritize estimated incremental contribution after all intervention costs. Where it is missing, use risk and value to select a pilot cohort rather than presenting an untested ranking as a profit forecast.
Expected net value per targeted customer = estimated additional probability of retention × contribution per additional retained customer, minus expected intervention cost.
Use a defined horizon for contribution, such as six months. Include expected service costs and remaining churn within that horizon. Subtract incentives, contacts, equipment and recovery costs consistently, without counting the same cost twice.
| Planning estimate | Segment A | Segment B |
|---|---|---|
| Churn risk before intervention | 60% | 30% |
| Additional retention probability | 1 percentage point | 4 percentage points |
| Contribution per additional retained customer | $34 | $84 |
| Expected cost per targeted customer | $1 | $1 |
| Expected net value per targeted customer | Negative $0.66 | Positive $2.36 |
The lower-risk segment has the better estimated campaign case. All figures above are invented planning examples, not customer results. The original rupiah examples were converted at about Rp17,900 per US dollar on 4 October 2026 and rounded to the nearest dollar, and the net values use those rounded inputs. The comparison assumes reliable estimates of additional retention; a churn score cannot supply those estimates by itself.
Build the workflow before buying a more complex model
- Define the outcome and clock. Specify what counts as churn and the prediction horizon. Distinguish suspension, prepaid inactivity, relocation and completed termination.
- Use information available at scoring time. Exclude later cancellation records and post-outcome notes. Otherwise the model can appear accurate by using information your team would not have had.
- Create an actionable queue. Each case needs a reason to investigate, an eligible response, an owner and a decision deadline.
- Record the full contact journey. Track assignment, attempted contact, reach, accepted offer, resolution and later account status separately.
- Evaluate outcomes against usual service. Compare similar eligible customers over the same horizon, preferably using randomized assignment for discretionary campaigns.
A simple repeat-complaint rule can be a useful starting baseline. Add a model when it improves useful prioritization on newer customer data and fits the team’s capacity. Check precision among the accounts agents can actually handle, recall of real churners and whether predicted probabilities match observed outcomes.
For FTTH providers operating across different cities or partner networks, inspect results by tenure, product and territory. Do not assume that a rule learned in an apartment cohort transfers unchanged to landed housing or a different prepaid model.
Calculate incremental retention campaign ROI
Compare churn among all customers assigned to the campaign with churn in a comparable usual-service group. Use the same follow-up window. The defaults show a hypothetical case, not market data or a prediction model.
Manual calculation: additional retained customers = assigned customers × (usual-service churn minus campaign churn) ÷ 100. Benefit = additional retained customers × contribution. Costs = assigned customers × average intervention cost + fixed costs. ROI = (benefit minus costs) ÷ costs × 100.
Include contact and expected incentive costs across all assigned customers, including those who accept an offer but would have stayed anyway. Set contribution over one explicit horizon, after normal service costs. Enter intervention costs separately. The model assumes a uniform contribution value, excludes financing and taxes, and does not estimate uncertainty or prove causality. Negative results are possible. Non-random comparisons can be biased.
A 5 percentage point difference means 50 additional retained customers per 1,000 assigned, not 150 “saves” simply because 150 people accepted an offer. Retained revenue is useful reporting, but contribution after costs determines the business case.
Give the CMO and sales leader a scorecard they can trust
| Metric | What it answers | Common reporting mistake |
|---|---|---|
| Contact and acceptance rates | Can the team reach customers and present a relevant offer? | Calling acceptance incremental retention |
| Verified recovery | Was the reported problem actually resolved? | Equating a closed ticket with a satisfied customer |
| Churn difference over the same horizon | Did the campaign improve retention against the comparison group? | Comparing contacted customers with a different population |
| Incremental contribution and ROI | Did the additional retained value exceed campaign costs? | Ignoring discounts given to customers who would have stayed |
Track retention beyond the offer expiry where practical, report cohort sizes and attach uncertainty estimates with analytics support. Random assignment improves causal interpretation, but small samples and unstable outcomes still require caution.
Use prediction responsibly: restrict subscriber-data access, minimize collection and keep human review for ambiguous cases. Low engagement is not a reason to deny support. A campaign can be selective while basic service quality remains a commitment to every customer.
A practical first 90 days
During the first month, define churn and the evaluation horizon, choose a fixable problem and establish a rule-based baseline. During the second, run one eligible intervention with a suitable comparison group and record costs and outcomes. During the third, review early contribution, recovery and workload, then continue follow-up for the full chosen horizon.
Do not claim a six-month retention result after a 90-day rollout. Scale when the economics, service capability and evidence support it. If the intervention shows no additional retention, improve the action before spending more on the prediction model.
Frequently asked questions
Can an ISP predict churn without machine learning?
Yes. Rules based on repeat complaints, failed payments or unresolved activation issues can identify accounts for investigation. Test them as a baseline and add more complex models only when they improve useful, actionable prioritization.
Is a high churn score enough to justify a discount?
No. Verify the problem, consider customer contribution and estimate the value of an effective action. A discount may not fix a fault, and the highest-risk customer may not be persuadable.
What is the difference between churn prediction and uplift?
Churn prediction estimates who may leave. Uplift estimates how an intervention changes retention compared with no additional intervention. Profitable targeting needs that incremental effect, not only a risk score or offer-acceptance probability.
What accuracy should a broadband churn model achieve?
There is no universal threshold. Evaluate on newer local data, compare with a baseline and examine precision, recall and probability calibration. Then test whether the actions informed by the model generate incremental contribution.
Evidence and methodology
Research checked on 3 October 2026. All campaign values and segment comparisons are illustrative. Signal mappings, workflow and rollout timing are editorial recommendations, not measured ISP results.
- Abdelhady and Mohamed, Scientific Reports: Leveraging artificial intelligence for predictive customer churn modeling in telecommunications, published 12 December 2025. Public-dataset model evaluation; its accuracy and recall are not evidence of FTTH campaign ROI.
- Verhelst and colleagues: A churn prediction dataset from the telecom sector, a new benchmark for uplift modeling, submitted 12 December 2023. Orange Belgium research benchmark presented in a workshop context. It informs the distinction between prediction and intervention, not an Indonesian retention forecast.