Forecasting Non-Performing Loans in Bangladesh: Evidence from ARIMA and Markov-Switching Autoregressive Models

Abstract

This study investigates the dynamics of non-performing loans (NPLs) across Bangladesh’s banking sector using quarterly data from 2007 to 2024. Employing Autoregressive Integrated Moving Average (ARIMA) and Markov-Switching Autoregressive (MSAR) models, we analyze NPL behavior across state-owned commercial banks, specialized banks, private commercial banks, foreign commercial banks, and industry totals. Structural break analysis using the PELT (L2) algorithm identifies critical shifts aligned with major economic shocks and policy changes, including the global financial crisis, regulatory reforms, COVID-19 moratorium, and recent classification threshold adjustments. Results reveal state-owned commercial banks and specialized banks contribute predominantly to rising aggregate NPLs, exhibiting persistent high-NPL regimes. While the MSAR models are effective at capturing the regime-switching NPL dynamics, the Auto-ARIMA specifications provide the best out-of sample validation for forecasting. Eight-quarter forecasts (2025-2026) using the optimized ARIMA model offer reliable and stable predictive accuracy compared to multi-regime alternatives. These findings provide critical insights for regulators and policymakers to enhance macro-prudential oversight and implement proactive credit risk monitoring strategies.