Prediksi Churn Nasabah Perbankan Menggunakan Metode Random Forest dan Logistic Regression Berbasis Optimasi Hyperparameter

Abstract

Churn nasabah perbankan merupakan kondisi ketika nasabah mengakhiri hubungan bisnis secara menyeluruh sehingga berpotensi mengancam stabilitas finansial perbankan. Oleh karena itu, diperlukan strategi prediksi churn nasabah yang akurat. Penelitian ini membandingkan kinerja dua metode analisis untuk memprediksi churn yaitu random forest dan logistic regression yang dioptimasi melalui hyperparameter tuning. Menggunakan 10,000 observasi nasabah dari platform Kaggle, strategi random undersampling dan cluster undersampling diterapkan sebagai data balancing. Evaluasi model menggunakan AUC, recall, precision, dan confusion matrix, sedangkan interpretasi fitur memanfaatkan feature importance, odds ratio, dan SHAP. Hasil menunjukkan random forest memiliki kinerja prediksi lebih baik dengan AUC 0.8579 pada random undersampling 70%, sementara logistic regression mencapai AUC 0.8334 pada random undersampling 30%. Skenario cluster undersampling terbukti paling efektif memaksimalkan recall. Analisis mengidentifikasi jumlah produk, usia, dan status keaktifan sebagai prediktor utama. Nasabah di atas 46 tahun dengan lebih dari dua produk namun pasif paling berisiko churn sehingga retensi pada segmen ini perlu diprioritaskan.

Customer churn in banking is a condition in which customers completely terminate their business relationships, potentially threatening the financial stability of banks. Therefore, an accurate customer churn prediction strategy is required. This study compares the performance of two analytical methods for predicting churn, namely random forest and logistic regression, optimized through hyperparameter tuning. Utilizing 10,000 customer observations from the Kaggle platform, random undersampling and cluster undersampling strategies were applied for data balancing. Model evaluation utilized AUC, recall, precision, and a confusion matrix, while feature interpretation employed feature importance, odds ratios, and SHAP. The results indicate that random forest exhibited better predictive performance with an AUC of 0.8579 at 70% random undersampling, while logistic regression achieved an AUC of 0.8334 at 30% random undersampling. The cluster undersampling scenario proved to be the most effective in maximizing recall. The analysis identified the number of products, age, and activity status as the primary predictors. Passive customers over 46 years old with more than two products are at the highest risk of churning; thus, retention efforts for this segment must be prioritized.