Device insurer achieves 90% churn prediction accuracy with ML-powered retention model

A major global device insurance provider was losing customers to both active policy cancellations and passive non-renewals — a dual churn problem that directly eroded renewal revenue and forced higher customer acquisition spend to compensate. In Property & Casualty, where policy lifetime value is the primary growth lever, undetected churn compounds quickly: customers acquired at significant cost exit silently, often without any retention attempt. The insurer had no mechanism to identify at-risk policyholders before they lapsed. Privacy regulations further constrained the problem, ruling out demographic data and limiting the team to behavioral and product signals — making early warning detection harder without the right analytical approach.

Tesseract Academy developed a two-stage machine learning pipeline tailored to the insurer's behavioral and product data. The first stage used classification models to produce a per-customer churn probability score — a numerical risk rating indicating likelihood of cancellation or non-renewal within a given window. The second stage applied survival analysis, a technique adapted from actuarial and medical research, to estimate the time-to-churn for each policyholder. Together, these produced two actionable outputs per customer: a risk score and a predicted churn date. The pipeline was trained on historical policy data including device type, tenure, geography, and usage patterns. Tesseract also delivered a ranked list of the top churn-driving features, giving the insurer interpretable business intelligence alongside the model outputs.

The model correctly identified approximately 89% of customers who would churn (churn detection rate), with prediction precision reaching up to 90% — meaning roughly nine out of ten flagged customers were genuine churn risks. This allowed the insurer to concentrate retention resources on the 15–20% of the customer base identified as high-risk, rather than applying costly blanket campaigns across the full book.

• 89% churn detection rate: the vast majority of actual churners were captured before they left
• Up to 90% precision: minimal false positives, protecting retention budget from wasted outreach
• Urgency-based prioritization: survival analysis enabled the team to rank outreach by predicted churn date, ensuring the highest-urgency cases were contacted first
• Targeted campaigns could deliver personalized incentives to flagged customers, shifting retention from reactive to proactive

• Frame the problem in business terms from the start — defining success as 'who will leave and when' (not model accuracy alone) keeps the project anchored to outcomes that retention teams can act on.
• Combine classification with survival analysis — a churn probability score tells you who is at risk; a time-to-churn estimate tells you when to act, which is essential for urgency-based prioritization and maximizing limited retention budgets.
• Behavioral data can be sufficient — privacy constraints that eliminate demographic signals are common in insurance; device, tenure, geography, and usage patterns alone can support a high-accuracy model.
• Churn prediction methods are portable — the classification-plus-survival-analysis framework applies to any subscription business with sufficient policy or behavioral history.