Bank Customer Repurchase Prediction: A Novel Weighted LightGBM

Customer repurchase frequency is a key indicator for measuring bank customer loyalty. To address the shortcomings of traditional customer management approaches in terms of efficiency and accuracy, as well as data imbalance issues, we aim to construct a customer repurchase predicition model based on LightGBM. We integrated de-identified data provided by the bank and successfully trained a three-class classification model through feature engineering (including behavioral feature aggregation and cross feature construction) and category weighting tailored to business requirements. The model achieved a weighted accuracy of 0.7747 and delivered a recall rate of 91.95% for high-frequency customers on the test set, which outperformed other gradient boosting methods and traditional models. The classification model we constructed provides technical support for the bank's refined customer management and precision marketing initiatives.