Estimasi Probabilitas Gagal Bayar Kredit Berbasis Tabnet Dengan Pendekatan Kalibrasi Bayesian

Salsabila, Farica (2026) Estimasi Probabilitas Gagal Bayar Kredit Berbasis Tabnet Dengan Pendekatan Kalibrasi Bayesian. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5002221061-Undergraduate_Thesis.pdf Restricted to Repository staff only Download (13MB) | Request a copy

Abstract

Estimasi probabilitas gagal bayar yang akurat dan reliabel merupakan komponen kritis dalam manajemen risiko kredit perbankan. Model deep learning modern, khususnya Attentive Interpretable Tabular Learning (TabNet), mampu menghasilkan kemampuan diskriminasi yang tinggi sekaligus menyediakan interpretabilitas intrinsik melalui mekanisme sequential attention. Namun, skor kepercayaan (confidence) yang dihasilkan cenderung overconfident sehingga tidak dapat digunakan langsung sebagai estimasi probabilitas gagal bayar yang reliabel tanpa tahap kalibrasi. Kondisi ini berpotensi menimbulkan kesalahan pengambilan keputusan dalam penilaian kelayakan kredit. Penelitian ini mengevaluasi efektivitas tiga metode kalibrasi post-hoc, yakni Bayesian Binning into Quantiles (BBQ), Temperature Scaling (TS), dan Isotonic Regression (IR) dalam meningkatkan reliabilitas estimasi probabilitas gagal bayar pada model TabNet menggunakan credit card default risk dataset dengan ketidakseimbangan kelas sebesar 8,12%. Model TabNet sebelum kalibrasi menghasilkan ROC AUC sebesar 99,36% namun menunjukkan gejala overconfidence yang signifikan dengan Expected Calibration Error (ECE) sebesar 4,68% dan Brier Score sebesar 3,68%. Setelah kalibrasi, BBQ memberikan perbaikan ECE terbesar 97,87% dengan nilai akhir ECE 0,0996%, diikuti IR sebesar 93,57%, dan TS hanya 0,51%. Kalibrasi BBQ juga meningkatkan precision sebesar 7,13% dan F1-Score sebesar 1,78% tanpa penurunan ROC AUC yang berarti. Mekanisme Bayesian model averaging pada BBQ terbukti mampu mengoreksi pola miskalibrasi secara adaptif pada seluruh rentang probabilitas, sehingga menghasilkan estimasi probabilitas gagal bayar yang lebih terkalibrasi dan dapat dipercaya. Hal ini menjadikan BBQ sebagai metode kalibrasi terbaik untuk estimasi probabilitas gagal bayar pada model TabNet dengan data kredit tidak seimbang.
=================================================================================================================================
Accurate and reliable estimation of default probability is a critical component of banking credit risk management. Modern deep learning models, particularly Attentive Interpretable Tabular Learning (TabNet), are capable of achieving high discriminatory capability while providing intrinsic interpretability through a sequential attention mechanism. However, the resulting confidence scores tend to be overconfident and therefore cannot be used directly as reliable estimates of default probability without a calibration stage. This condition has the potential to lead to decision-making errors in creditworthiness assessments. This study evaluates the effectiveness of three post-hoc calibration methods, namely Bayesian Binning into Quantiles (BBQ), Temperature Scaling (TS), and Isotonic Regression (IR), in improving the reliability of default probability estimates in the TabNet model using a credit card default risk dataset with a class imbalance of 8.12%. The TabNet model before calibration produced an ROC AUC of 99.36%, but showed significant overconfidence, with an Expected Calibration Error (ECE) of 4.68% and a Brier Score of 3.68%. After calibration, BBQ provided the largest ECE improvement of 97.87%, with a final ECE of 0.0996%, followed by IR at 93.57%, and TS at only 0.51%. BBQ calibration also increased precision by 7.13% and the F1-Score by 1.78% without a significant decrease in ROC AUC. The Bayesian model averaging mechanism in BBQ has been shown to adaptively correct miscalibration patterns across the entire probability range, resulting in more calibrated and reliable default probability estimates. This makes BBQ the best calibration method for estimating default probability in TabNet models with imbalanced credit data.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Bayesian Binning into Quantiles, Expected Calibration Error, Kalibrasi Probabilitas, Probabilitas Gagal Bayar, TabNet, Bayesian Binning into Quantiles, Expected Calibration Error, Probability Calibration, Probability of Default, TabNet.
Subjects:	Q Science > QA Mathematics > QA279.5 Bayesian statistical decision theory.
Divisions:	Faculty of Mathematics, Computation, and Data Science > Mathematics > 44201-(S1) Undergraduate Thesis
Depositing User:	Farica Salsabila
Date Deposited:	28 Jul 2026 02:28
Last Modified:	28 Jul 2026 02:31
URI:	http://repository.its.ac.id/id/eprint/137990

Actions (login required)

View Item