Recency windows are a staple of churn modelling, but they are often chosen heuristically. Following the launch of “VIP 2.0” at Sky, we needed to reassess how long customer propositions reduce churn risk, replacing legacy window assumptions with a data-driven approach.

The challenge was combinatorial: 11 propositions, alternative starting points (0 vs 1 month), and window lengths from 1 to 24 months produced a vast search space of possible configurations. Rather than exhaustively testing every combination, we used Latin hypercube sampling to efficiently explore the space, generating 10,000 candidate configurations. A matched-pairs churn framework, already well understood by stakeholders, was re-implemented in R using data.table for speed, reducing iteration time from minutes in BigQuery to seconds locally.

Convergence was typically achieved within a few thousand iterations.Two complementary analyses were then applied. First, marginal distributions were constructed by varying one proposition at a time to estimate its standalone effect. Second, to capture interaction effects between propositions, a gradient boosted tree was trained on the full sample of configurations to predict total accounts saved. Shapley values (via shapviz) were used to decompose model predictions, allocating the incremental churn benefit across propositions and window lengths in an additive and interaction-aware way.

Encouragingly, both approaches aligned closely. The SHAP analysis provided a robust validation of the marginal method and highlighted clear decay patterns, in some cases showing sharp drop-offs in effectiveness beyond specific horizons. The final outputs informed conservative, data-backed recency windows, now embedded within our GCP reporting pipeline.This session will cover the sampling strategy, high-performance R implementation, and how SHAP values enabled practical, interpretable decision-making in a complex retention programme.