Interfacing learning methods for anomaly detection in multi-country financial stress indicators

Abstract

This paper presents a novel ensemble supervised learning classification model designed for the early detection of financial stability anomalies. In particular, we utilise the time series of Financial Stress Indices (FSI) across multiple countries in developing an early-warning system. The innovation of this model lies in its unique integration of stochastic process modelling, hidden Markov models (HMM), random forest (RF), and XGBoost algorithms. This results to a comprehensive approach that can capture the dynamics of FSIs and forecast potential crisis episodes. The model's strength arises from the synthesis of the Ornstein–Uhlenbeck (OU) processes and HMM online recursive filters, forming a robust framework. Additionally, a feature selection module based on RF and a final classifier using XGBoost enhance the out-of-sample predictive performance. Our comparative analyses with five alternative models underscore the strong predictive power of the proposed model. A tailored feature-importance analysis highlights the substantial impact of the HMM features, emphasising their crucial role in the model's effectiveness. Furthermore, the inclusion of two projected anomaly-warning signals enhances the model's ability to predict extreme events, benefitting financial stability and public policy research. © 2024 The Author(s)