An early-warning system for emerging market crises: a machine-learning approach

Thesis / Dissertation

2026

Permanent link to this Item
Authors
Supervisors
Journal Title
Link to Journal
Journal ISSN
Volume Title
Publisher
Publisher

University of Cape Town

License
Series
Abstract
This research report develops an interpretable machine-learning early-warning system for financial crises in emerging market economies. While traditional early-warning system models identify key macro-financial predictors, their reliance on linear specifications often fails to capture the non-linear dynamics and cross-country heterogeneity inherent in modern crises. In contrast, recent machine-learning approaches offer greater predictive flexibility, yet are frequently criticised for lacking policy relevance due to their “black-box” opacity. Although the trigger events surrounding a crisis can materialise rapidly, the predictive content of crises lies primarily in the gradual accumulation of macro-financial vulnerabilities over multiple years, a process that is well captured at quarterly frequency and that aligns with the medium-term horizon over which macroprudential policy operates. Using a quarterly panel of 25 emerging market economies from 1995 to 2024, this study constructs indicators for currency, banking, and sovereign debt crises. Gradient-boosting models are estimated within a walk-forward expanding-window framework to generate probabilistic forecasts at four- and twelve-quarter horizons. Performance is assessed using discrimination, calibration, and policy-usefulness metrics that reflect the asymmetric costs of missed crises relative to false alarms. The results demonstrate that ensemble methods, particularly gradient-boosting classifiers, consistently outperform linear benchmarks, with predictive power improving consistently at longer horizons, supporting the focus on medium-term horizons for policy surveillance. Shapley Additive exPlanations (SHAP)-based interpretability reveals that crisis risk is shaped by non-linear threshold effects in domestic credit conditions, external imbalances, and global financial conditions. Overall, the findings suggest that machine-learning frameworks can significantly enhance macroprudential surveillance by combining improved predictive accuracy with the transparency required for timely policy intervention.
Description

Reference:

Collections