Pengembangan Model Prediksi Persetujuan Pinjaman Menggunakan Arsitektur Hybrid Stacking Ensemble Berbasis TabPFN

Ariatama, Ilham Putra (2026) Pengembangan Model Prediksi Persetujuan Pinjaman Menggunakan Arsitektur Hybrid Stacking Ensemble Berbasis TabPFN. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6025241001-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (3MB) | Request a copy

Abstract

Industri perbankan berperan strategis dalam mendukung pertumbuhan ekonomi global melalui penyaluran kredit, namun menghadapi tantangan dalam prediksi persetujuan pinjaman yang akurat untuk meminimalkan risiko kredit. Metode penilaian kredit tradisional yang mengandalkan analisis manual memiliki keterbatasan dalam efisiensi dan konsistensi. Meskipun machine learning telah diterapkan secara luas, permasalahan class imbalance, heterogenitas fitur, dan keseimbangan antara akurasi dengan interpretabilitas model masih menjadi tantangan fundamental. Penelitian ini bertujuan mengembangkan dan mengoptimalkan model prediksi persetujuan pinjaman menggunakan arsitektur hybrid stacking ensemble yang mengintegrasikan TabPFN (Tabular Prior-Data Fitted Network) sebagai meta-learner dengan Logistic Regression, Random Forest, dan XGBoost sebagai base learners. Penelitian ini menggunakan tiga dataset publik dengan karakteristik berbeda, yaitu Loan Prediction Problem (614 sampel), German Credit Card Dataset (1.000 sampel), dan Loan Approval Classification Dataset (45.000 sampel). Metodologi meliputi data preprocessing, feature engineering, dan stratified train-test split (80:20), dengan evaluasi menggunakan confusion matrix, Accuracy, Precision, Recall, dan F1-Score pada tujuh konfigurasi stacking ensemble. Hasil eksperimen menunjukkan bahwa arsitektur hybrid stacking ensemble yang diusulkan secara konsisten meningkatkan akurasi dibandingkan model klasifikasi tunggal pada ketiga dataset. TabPFN dengan data original menjadi single classifier terbaik pada seluruh dataset (akurasi 82,29%; 77,50%; dan 92,90%). Stacking ensemble dengan TabPFN sebagai meta-learner meningkatkan akurasi lebih lanjut menjadi 84,38%; 80,00%; dan 93,07%, atau peningkatan 0,17%–2,50% poin persentase dibandingkan single classifier terbaik. Perbandingan dengan tiga konfigurasi stacking konvensional lain sebagai meta-learner menunjukkan tidak satu pun mampu melampaui akurasi TabPFN sebagai single classifier, membuktikan bahwa peningkatan performa berasal secara spesifik dari kapasitas TabPFN sebagai meta-learner, bukan sekadar efek generik arsitektur stacking. Konfigurasi base learner optimal bervariasi sesuai skala dan heterogenitas dataset. Kombinasi Logistic Regression dan XGBoost optimal untuk dataset kecil-menengah, sementara kombinasi lengkap ketiga base learner optimal pada dataset berskala besar dan lebih heterogen. Data original tanpa resampling secara konsisten menghasilkan akurasi terbaik, mengindikasikan ketahanan bawaan arsitektur ini terhadap ketidakseimbangan kelas. Penelitian ini berkontribusi secara teoritis dalam membuktikan efektivitas TabPFN sebagai meta-learner dalam kerangka stacking ensemble untuk data tabular kredit, sekaligus menyediakan framework bagi lembaga keuangan untuk meningkatkan akurasi credit scoring tanpa kompleksitas hyperparameter tuning yang ekstensif.
==========================================================================================================================================
The banking industry plays a strategic role in supporting global economic growth through credit disbursement yet faces challenges in accurately predicting loan approvals to minimize credit risk. Traditional credit assessment methods relying on manual analysis have limited efficiency and consistency. Although machine learning has been widely applied, class imbalance, feature heterogeneity, and the trade-off between Accuracy and interpretability remain fundamental challenges. This study develops and optimizes a loan approval prediction model using a hybrid stacking ensemble architecture that integrates TabPFN (Tabular Prior-Data Fitted Network) as meta-learner with Logistic Regression, Random Forest, and XGBoost as base learners. This study uses three public datasets with different characteristics: the Loan Prediction Problem (614 samples), German Credit Card Dataset (1,000 samples), and Loan Approval Classification Dataset (45,000 samples). The methodology includes data preprocessing, feature engineering, and a stratified train-test split (80:20), with evaluation using confusion matrix, Accuracy, Precision, Recall, and F1-Score across seven stacking ensemble configurations. Experimental results show that the proposed hybrid stacking ensemble consistently improves Accuracy over single classification models across all three datasets. TabPFN with original data was the best single classifier across all datasets (Accuracy of 82,29%, 77,50%, and 92,90%). Stacking with TabPFN as meta-learner further improved Accuracy to 84,38%, 80,00%, and 93,07%, a gain of 0,17%–2,50% percentage points over the best single classifier. Comparison with three conventional stacking configurations using other classifiers as meta-learners showed nonsurpassed TabPFN's Accuracy as single classifier, confirming that the improvement stems specifically from TabPFN's capacity as meta-learner rather than a generic stacking effect. The optimal base learner configuration varied by dataset scale and heterogeneity: Logistic Regression combined with XGBoost was optimal for small-to-medium datasets, while the full three-learner combination was optimal for the larger, more heterogeneous dataset. Original data without resampling consistently produced the best Accuracy, indicating the architecture's inherent robustness to class imbalance. This study contributes theoretically by demonstrating TabPFN's effectiveness as a meta-learner within a stacking ensemble framework for tabular credit data, while also providing a framework for financial institutions to improve credit scoring Accuracy without extensive hyperparameter tuning.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	class imbalance, credit scoring, hybrid stacking ensemble, prediksi persetujuan pinjaman, loan approval prediction, TabPFN class imbalance, credit scoring, hybrid stacking ensemble, loan approval prediction, TabPFN
Subjects:	H Social Sciences > HG Finance H Social Sciences > HG Finance > HG3751 Credit--Management. T Technology > T Technology (General) T Technology > T Technology (General) > T57.5 Data Processing T Technology > T Technology (General) > T58.6 Management information systems
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55101-(S2) Master Thesis
Depositing User:	Ilham Putra Ariatama
Date Deposited:	24 Jul 2026 21:24
Last Modified:	24 Jul 2026 21:24
URI:	http://repository.its.ac.id/id/eprint/137311

Actions (login required)

View Item