What Actually Drives Profit in Telecom Churn Retention: Profit-Aligned Targeting versus Accuracy and Calibration

This paper presents a controlled, significance-tested decomposition of a telecom customer-churn retention pipeline, isolating which components actually improve realised profit rather than accuracy. Using two independent telecom datasets and bootstrap confidence intervals, we show that aligning the targeting decision with profit yields a large, statistically significant gain over accuracy-optimal thresholding (+13.1 percentage points of profit capture on Maven Telecom), while widely used techniques—class re-sampling and probability calibration—add no statistically significant profit once this is done. We further show that re-sampling harms probability calibration without improving ranking, and we honestly delimit the conditions under which profit-aligned targeting does and does not help.