Predicting abnormal capital flow episodes with machine learning methods

In recent years, public health emergencies and geopolitical conflicts have constantly triggered volatility in the global economy and financial markets. Such frequent shocks have led to abnormal capital flow episodes, which can destabilize financial systems and foreign exchange markets, and sometimes these episodes are precursors to financial crises. Therefore, we develop an early warning model for abnormal capital flow episodes with a forecast horizon set two quarters in advance, employing two traditional linear regression models and nine machine learning algorithms. We also utilize two ensemble technologies, voting and stacking, to enhance out-of-sample predictive accuracy. This provides monetary authorities across nations with a practical early warning model that allows manual control over the forecast horizon and delivers robust predictive performance on out-of-sample observations, enabling timely interventions and preventative measures against risks associated with volatile capital flows. Furthermore, causal analysis using Shapley value decomposition and Shapley regression reveal drivers and mechanisms of abnormal capital flow episodes that differ from those identified by traditional linear models. For instance, the Shapley-based interpretation uncovers complex nonlinear relationships and highlights previously overlooked variables, such as domestic liability dollarization, as crucial predictors of sudden stops and capital flight. The Shapley-based interpretation reveals that the relative importance of predictors shifts after the 2008 Global Financial Crisis: features such as DLD become far more influential in the post-GFC period, reflecting a transition in investor behavior from profit-seeking to risk-averse. This insight deepens our understanding of the complex dynamics influencing international capital movements and enhances risk management tools in an interconnected world.