Prediksi Risiko Default Kredit Usaha Rakyat Berjalan Berbasis Snapshot Time dan Pelabelan Forward-Looking

Jaya, Muhammad Triyanda Taruna (2026) Prediksi Risiko Default Kredit Usaha Rakyat Berjalan Berbasis Snapshot Time dan Pelabelan Forward-Looking. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6025241014-Master_Thesis.pdf Restricted to Repository staff only Download (3MB) | Request a copy

Abstract

Penyaluran Kredit Usaha Rakyat (KUR) berperan penting dalam memperluas akses pembiayaan bagi UMKM, namun di sisi lain meningkatkan eksposur risiko default (gagal bayar) yang dapat menekan kualitas aset perbankan. Praktik penilaian risiko default yang lazim digunakan masih berfokus pada tahap pengajuan kredit (application scoring), sehingga belum optimal untuk memantau risiko pada kredit yang sedang berjalan (on-book). Penelitian ini mengembangkan model prediksi risiko default KUR berjalan dengan memanfaatkan kerangka snapshot time dan pelabelan forward-looking tiga bulan ke depan berdasarkan definisi default Otoritas Jasa Keuangan. Data yang digunakan berasal dari transaksi riil pembayaran angsuran KUR pada salah satu bank pembangunan daerah di Indonesia dan ditransformasikan menjadi 21.416 snapshot pinjaman dengan proporsi sekitar 8% default dan 92% non-default. Model dibangun menggunakan algoritma XGBoost, LightGBM dan Multilayer Perceptron (MLP) dengan berbagai strategi penanganan class imbalance (tanpa penyesuaian, SMOTE, SMOTE-ENN dan class weight), serta varian ensemble (hard voting, soft voting dan stacking). Hasil eksperimen menunjukkan bahwa rekayasa data perilaku pembayaran angsuran melalui kerangka snapshot time dan pelabelan forward-looking mampu memprediksi default secara efektif, dengan konfigurasi terbaik diperoleh pada MLP dengan SMOTE dan penentuan ambang keputusan berbasis F1 maksimum, yang pada data uji tahun 2024 menghasilkan AUC ≈ 0,996, F1 ≈ 0,83, serta kombinasi precision dan recall yang seimbang pada kelas default. Interpretasi model menggunakan SHAP pada tingkat global dan lokal mengonfirmasi bahwa indikator perilaku pembayaran—terutama paid_full_rate, avg_delay_flag dan rasio pembayaran—menjadi penentu utama skor risiko dan penjelasan model dinilai wajar dan dapat diterima oleh analis kredit. Temuan ini menunjukkan bahwa pendekatan snapshot time dan pelabelan forward-looking pada data KUR berjalan, yang masih jarang dibahas dalam konteks Indonesia, dapat menghasilkan model prediksi yang akurat dan dapat dijelaskan.
============================================================================================================================================
The distribution of micro, small and medium enterprise (MSME) loan program that backed by the goverment plays a key role in expanding financing access for MSMEs, but at the same time increases exposure to default risk that can undermine banks’ asset quality. Conventional credit risk assessment practices still focus on the loan application stage (application scoring), and therefore are not yet optimal for monitoring risk on loans that are already on the books (on-book). This study develops a default risk prediction model for active KUR loans by employing a snapshot time framework and three-month forward-looking labeling based on the non-performing definition of the Financial Services Authority of Indonesia. The data are sourced from real KUR installment payment transactions at a regional development bank in Indonesia and transformed into 21,416 loan snapshots with a class distribution of approximately 8% default and 92% non-default. The models are built using XGBoost, LightGBM, and Multilayer Perceptron (MLP) algorithms with several class imbalance handling strategies (no adjustment, SMOTE, SMOTEENN, and class weighting), as well as ensemble variants (hard voting, soft voting, and stacking). The experimental results show that engineering installment payment behavior using the snapshot time framework and forward-looking labeling can effectively predict loan default, with the best configuration obtained from MLP with SMOTE and an F1-maximizing decision threshold, yielding AUC ≈ 0.996, F1 ≈ 0.83, and a balanced combination of precision and recall for the default class on the 2024 test set. Model interpretation using SHAP at both global and local levels confirms that payment behavior indicators—particularly paid_full_rate, avg_delay_flag, and payment ratio features—are the main drivers of the risk score, and that the explanations are considered reasonable and acceptable by credit analysts. These findings indicate that the snapshot time and forward-looking labeling approach on active KUR data, which has rarely been explored in the Indonesian context, can produce prediction models that are both accurate and explainable.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Forward-Looking Label, Kredit Usaha Rakyat, Machine Learning, Prediksi Default, Snapshot Time, Default Prediction, Forward-Looking Label, Machine Learning, MSME, Snapshot Time
Subjects:	T Technology > T Technology (General) > T174 Technological forecasting T Technology > T Technology (General) > T58.5 Information technology. IT--Auditing
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55101-(S2) Master Thesis
Depositing User:	Muhammad Triyanda Taruna Jaya
Date Deposited:	02 Feb 2026 01:02
Last Modified:	02 Feb 2026 01:02
URI:	http://repository.its.ac.id/id/eprint/131435

Actions (login required)

View Item