금융시장 위험 회피 국면 예측을 위한 머신러닝 모델

Machine Learning Model for Predicting Risk-off States in Financial Markets

Purpose – This study develops a machine learning model to predict global financial market risk-off phases through the analysis of carry trade dynamics. Design/Methodology/Approach – Because carry trade activities are difficult to observe directly, we construct three principal indicators via Principal Component Analysis (PCA): Profitability (ECP), Liquidity (ECL), and Risk (ECR). ECP measures risk-adjusted returns using interest rate differentials and exchange rate volatility. ECL captures execution feasibility and liquidity based on long- and short-term interest spreads, the dollar index, global liquidity, and central bank assets. ECR quantifies risk aversion through volatility indices such as the VIX and S&P500 realized volatility. Findings – Using these indicators with a Random Forest algorithm, we estimate the probability of risk-off events. The model was trained on data from March 2003 to December 2019 and tested on post-2020 data. Results demonstrate strong predictive performance, with an Average Precision of 0.84 from the Precision-Recall curve. Variable importance analysis highlights ECR and ECL as dominant predictors, indicating that heightened risk aversion and reduced liquidity—often linked to carry trade unwinding—are primary channels of instability. Extending the model with external variables, including Fed rates, global supply chain pressures, and oil prices, further improved accuracy. Importantly, true positive predictions aligned with substantial equity downturns, averaging a -152.3% cumulative KOSPI decline, confirming the model’s ability to anticipate market stress. Research Implications – In conclusion, this Random Forest-based prediction model, utilizing comprehensive carry trade-related indicators, offers early detection of financial market shifts and valuable insights for investment decisions.