Hybrid early warning system: Integration of Z-score and machine learning for predicting financial performance of IRB in Indonesia

This study aims to develop a hybrid Early Warning System (EWS) that integrates the traditional Z-score method with machine learning algorithms to accurately predict the financial performance and potential failure of IRB in Indonesia. The study employs a data science–driven approach using monthly financial and macroeconomic data from 2011 to 2024. A total of 21 independent variables were tested against the Z-score as the dependent variable. Various machine learning models were compared using AutoML (TPOT and LazyRegressor), and model validation was conducted through statistical testing, cross-validation, robustness checks, and feature importance analysis. The Extra Trees algorithm emerged as the most accurate and robust predictive model, achieving an R² of 0.95 and the lowest Mean Absolute Error (MAE) of 0.0676. Key predictive variables identified include Non-Performing Financing (NPF), inflation (INF), and istishna'-based Financing. The model successfully anticipates financial distress up to two years in advance and enables the categorisation of IRB conditions into Stable, Vulnerable, and Unstable using Z-score thresholds. This is the first study to integrate Z-score with advanced machine learning techniques within a hybrid EWS framework tailored to IRB. It offers a novel predictive model that enhances early risk detection and supports regulatory decision-making. The study focuses exclusively on IRB in Indonesia and may require recalibration for broader application in different financial systems or geographies. Reporting accuracy and completeness may also limit the use of secondary data. The hybrid model provides actionable insights for regulators (OJK, LPS) and financial institutions to monitor IRB's health proactively. It strengthens risk mitigation strategies and contributes to the long-term resilience of the Islamic microfinance sector.