Customer Churn Prediction on Structured Data Using FT-Transformer and Stacking Ensembles

Researchers propose a hybrid machine learning architecture combining FT-Transformer neural networks with XGBoost gradient boosting to predict customer churn in banking and subscription services. The ensemble method achieves superior performance metrics (62.10% F1, 0.861 AUC-ROC) compared to baseline models while addressing critical challenges in class imbalance and probability calibration.

This research addresses a practical yet underexplored problem in machine learning: combining modern deep learning with classical ensemble methods for structured tabular data. Customer churn prediction directly impacts profitability across financial services, telecom, and SaaS industries, where retention costs significantly less than acquisition. The study's hybrid approach recognizes that no single algorithm dominates tabular prediction tasks—transformers excel at capturing complex nonlinear interactions while tree-based methods provide robust decision boundaries with minimal hyperparameter tuning.

The methodology reflects broader trends in machine learning engineering: moving beyond single-model approaches toward carefully calibrated ensembles. The use of out-of-fold stacking with logistic regression meta-learning demonstrates maturity in handling the bias-variance tradeoff. Critically, the authors address reproducibility gaps endemic to prior churn research by implementing rigorous statistical validation with confidence intervals and ablation studies. This contrasts sharply with much published ML work that lacks proper cross-validation or confidence reporting.

For financial institutions, this framework offers immediate practical value. The 3.37-point F1 improvement over MLP baselines translates to measurable gains in identifying at-risk customers, enabling targeted retention campaigns. The approach sidesteps synthetic oversampling techniques that can distort minority-class distributions, instead using class-weighted loss—a more principled solution. Banking and subscription platforms can implement this architecture with existing tools (XGBoost, scikit-learn, Hugging Face transformers).

Future development should evaluate performance on domain-specific churn datasets where feature distributions differ from public benchmarks. Integration with real-time prediction pipelines and exploration of additional meta-learner architectures remain open questions.

• →FT-Transformer and XGBoost ensemble achieves 62.10% F1 and 0.861 AUC-ROC on bank churn data, outperforming MLP baselines by 3.37 F1 points.
• →Out-of-fold stacking with logistic regression meta-learner recalibrates overconfident base models while learning optimal combination weights.
• →Class-weighted loss functions address imbalance without synthetic oversampling, preserving authentic minority-class distributions.
• →Rigorous statistical validation with 5x5 cross-validation and 95% confidence intervals strengthens reproducibility compared to prior churn prediction studies.
• →Hybrid architecture offers extensible reference implementation for tabular prediction tasks across banking, insurance, and SaaS platforms.