Skewness-aware Boosting Regression Trees for Customer Contribution Prediction in Financial Precision Marketing

In an era of digital evolution, banking sectors face the dual challenge of nurturing a digitally savvy demographic and managing potential dormant account holders. This study delves deep into the prediction of customer contributions, particularly considering the skewed nature of such data. The inherent skewness in customer contribution data, highlighted by the substantial low-value contribution group and the vast variability among high-value contributors, necessitates an advanced prediction model. Addressing this, we present the Skewness-aware Boosting Regression Trees (SBRT) framework to predict customer contributions whose distribution exhibit high skewness. SBRT seamlessly combines the strength of Gradient Boosted Decision Trees with a novel mechanism of random tree deactivation, adeptly tackling distribution skewness. The model’s effectiveness is rooted in four principles: cross-feature extraction, a percentile-based calibration and rebalancing method, tree deactivation during the boosting phase, and the utilization of Huber loss. Extensive real-world bank data testing underscores SBRT’s promising capability in managing skewed distributions, setting it a cut above in predicting customer contributions. The culmination of this work lies in its practical validation, where online A/B tests highlight SBRT’s tangible industrial applicability.