Churn prediction in a low-ARPU MNP market

Three ways to see it

1. 1

   Pakistan is the world's fifth-largest mobile market by subscribers and one of the lowest by ARPU, in the range of USD 1.40 per month. Margins are thin and acquisition costs are high. The Mobile Number Portability framework, live since 2007 under PTA, makes switching almost frictionless. The result is a market where retention beats acquisition by a wide margin, and where churn prediction is not a back-office model but a front-line revenue lever.

2. 2

   Way one to think about churn: it is not one event, it is a slope. A subscriber does not wake up one morning and decide to port. He decreases recharges over weeks. He stops topping up before the weekend. He drops out of one WhatsApp group, then another. He starts using more Wi-Fi at home. He uses competitor apps when on his other SIM. By the time he visits a franchise to port, he has been gone for a month. A model that watches the slope spots him at week one and intervenes at week two.

3. 3

   Way two: Jazz churn drivers and Zong churn drivers are not the same. Jazz subscribers tend to be older, more likely to use JazzCash, more likely to be primary-SIM holders. Their churn signal is heavily about voice quality and JazzCash reliability. Zong subscribers skew younger, gaming-heavy, data-bundle sensitive. Their churn signal is about throughput, latency, and competitive offers from Telenor or Ufone. A churn model trained on Jazz data does not transfer cleanly to Zong, and vice versa. Each operator must build a model that knows its own customer.

The why-tree

Try this with Claude

AI-edge prompt: 'I work in retention at a Pakistani mobile operator. Design a churn prediction pipeline using 90 days of recharge, voice MOU, data MB, complaint, and JazzCash transaction data, with a 30-day prediction window. Recommend three model families, three calibration techniques, and the SHAP-style features to expose to the retention agent. Flag any PDP Act 2025 risks I should ask my legal team about.'