Early Detection of Fraud Indications in Ultra-Micro Financial Institutions Using Machine-Learning-based Financial Journal Data Analysis

Abstract

Ultra Micro Financial Institutions (UMFI) play an important role in expanding financial inclusion, but their high transaction volume, relatively small transaction values, direct cash handling, and limited internal control capacity may create fraud risks that are difficult to detect from aggregated financial statements. This study develops an early fraud-indication detection framework using transaction-level accounting journal data from PT Permodalan Nasional Madani (PNM), Magelang Branch. The study adopts a quantitative predictive case-study and proof-of-concept design. The available dataset covers 1 January to 31 March 2026 and contains 305,343 transactions from five units, 104 unique general-ledger codes, and balanced debit and credit totals of IDR 160,274,807,743. The novelty lies in translating the Fraud Triangle—pressure, opportunity, and rationalization—into operational transaction features, including end-period activity, transaction spikes, outliers, duplication, unusual account combinations, reversals, and correction-related descriptions. Rule-based screening is used only to identify risky candidates, while final fraud and non-fraud labels must be validated by internal audit. The analytical pipeline includes anonymization, leakage-aware preprocessing, feature engineering, train-validation-test splitting, SMOTE only within training data, Stratified 5-Fold Cross-Validation, hyperparameter tuning, Logistic Regression, Decision Tree, Random Forest, threshold selection, confusion-matrix evaluation, ROC-AUC, PR-AUC, and SHAP interpretation. At the time of article preparation, validated labels and final experiments were not yet available; therefore, model performance and the best-performing model are not reported. The article focuses on the methodological framework, theory-based feature design, and a risk-scoring early warning system for internal audit support