Predicting customer churn is a critical task for businesses aiming to retain customers and maintain profitability. This research adopts an individual participant data metaanalysis (IPD-MA) approach to evaluate the effectiveness of various machine learning models in predicting customer churn across multiple publicly available datasets. This methodology facilitates a robust comparison and validation of predictive models by integrating raw data from different studies. The study employs a two-stage approach: first, individual datasets are analyzed to obtain machine learning performance metrics; second,
these aggregated metrics are combined using fixed-effect and random-effect meta-analysis models. The results reveal significant variability in model performance across different datasets, with ensemble methods like Catboost, Lightgbm, and Gradient Boosting consistently outperforming other models, achieving the highest average AUCs of 0.9036, 0.9000, and 0.8936, respectively. The study also highlights the importance of considering dataset-specific characteristics and model capabilities, as well as the necessity of accounting for heterogeneity in meta-analyses. This research makes several key contributions, including methodological advancements in applying IPD-MA to machine learning, and a
comprehensive evaluation of model performance. The findings offer a valuable reference for selecting and optimizing machine learning models in various industrial applications, guiding future research and practical implementations.
