Machine learning modeling to forecast uncertainty between capital sudden stop and boom

Abstract

In emerging economies, sudden stops of capital inflows boost the collapse of stock and exchange rate markets and plunge countries into unemployment, loss of production, and diminishing exports, leading to a financial crisis. Recently, the nonlinear relationship-based machine learning (ML) model for analyzing the complexity and uncertainty of financial and economic systems has been in the spotlight, but they are still poorly used for predicting sudden stops of capital. Because there is no verified tool that elaborately measures the indicia of rapid suspension or expansion of capital for prediction purposes, ML models learned from domains with inadequately defined economic characteristics of capital movements can suffer from poor predictive power and reliability. In addition, many economists do not trust the ML model due to the lack of interpretability caused by the black box structure of ML models. In this study, three approaches are proposed for better prediction and decision-making. First, using data for 37 emerging economies from 1990 to 2019, we apply various ML techniques such as extreme gradient boost (XGB), which is well known for the latest ensemble learning technology. Second, we analyze the causal relationship to the outcomes of our models using SHAP (SHapley Additive exPlanations) methods, a powerful technique for ML interpretation. Particularly, from the perspective of post-COVID-19, our model predicts an increased probability of sudden stops in countries where a sharp decline in real interest rates and exports is evident. Finally, we propose a tool to measure capital flows and assess excessive levels for forecasting purposes. Particularly, our tool extracts credit boom events that are highly correlated with sudden stops, and the ML models with credit boom events have significantly improved predictive performance. Specifically, in the forecast of a sudden stop after one year, the prediction accuracy of the models with a credit boom event is improved by 11.2% on average compared to models without credit boom information. In addition, the gap between the predicted hit and the miss rate in the proposed models was reduced to −16.2% on average compared to the original, improving the balance of classification.