The challenge
The personal-banker network of a retail financial services provider is its most valuable sales asset. The company had no quantitative tool to identify agents likely to leave within a 3–9 month horizon, so retention actions came too late.
At a glance
Data6 DWH tables, ~40,000 observations, 2 years
MethodsGBT, L2 logistic regression, 5-fold CV
DeliveryJoint with a Big Four advisory firm
What we did
SWICON, delivering jointly with a Big Four advisory firm, built a churn model on two years of data-warehouse history covering 5,708 agents and around 40,000 training observations. The label was defined precisely: agents who will produce zero sales for three consecutive months within the forward window, not a vague attrition risk. Six warehouse tables were joined, six-month lookback features and OLS trend features were engineered, and gradient-boosted trees were benchmarked against a regularised logistic regression as an interpretable alternative.
The result
The best model (GBT) reached AuC 0.79 with 76% recall, identifying 76 of every 100 departing agents 3–9 months before departure. Feature importance was interpreted across deposit contracts, cross-sell, sales performance and certifications. One insight stood out: campaign-like over-performance appeared as an early indicator of subsequent departure. This quarter's top performer can be next quarter's leaver.
