Predicting & Preventing Digital Disengagement

In a single-SIM market, silence is close to churn: users keep one line for years and switching is inconvenient. In much of West Africa most subscribers carry two or three SIMs and move between them daily, chasing promotions, bundles and coverage.

That breaks the usual churn definition. A customer can be silent for weeks and come back, or be silent for weeks because they have already gone. The operator usually notices only after revenue has fallen — which is too late to intervene.

• Framed the target as digital silence — no calls, SMS, data or recharge — rather than a formal account closure.
• Engineered month-on-month change features (data, recharge, calls, SMS, spend) so the model reads trajectory, not just level. A customer whose usage is halving is a different risk from one who has always been light.
• Added tenure, device class and registration channel as profile-side predictors alongside usage and financial behaviour.
• Removed non-subscriber lines (routers, modems, IoT) and non-behavioural usage categories before modelling, and excluded fields whose quality could not support them — a placeholder-heavy date of birth field among them.
• Trained logistic regression as the primary model for interpretability and odds ratios, with a CART decision tree for operational segmentation, on a held-out validation split.

Both models classified well overall, but they were not equivalent for the decision at hand. The tree was marginally better at confirming that an active customer would stay active; the logistic model was better at catching customers who were about to go silent. For a retention team, those are not the same thing, and the second one is the one worth optimising. Selecting on sensitivity rather than headline accuracy is the substantive choice in the project.

The tree earned its place elsewhere: it turns the same signal into if-then segments a campaign team can act on without reading a coefficient table.

Client-specific figures, dataset details and segment-level findings are withheld here. The methodology, the modelling reasoning and the selection logic are the transferable part.

• Two consecutive months of history is thin for a trajectory model. A longer window would separate seasonal dips from genuine decay.
• Silence is a censored outcome, not a clean label — survival modelling fits the question better than binary classification.
• The registration-quality signal points at a data-collection fix, not just a model feature. Fixing intake at the SIM point improves every downstream model.