금과 비트코인의 안전자산 효과: 머신러닝 기반 충격 탐지와 분위수 분석

Safe-Haven Analysis of Gold and Bitcoin: A Machine Learning and Quantile Regression Approach

Purpose – This study examines the safe-haven properties of gold and Bitcoin by analyzing their asymmetric responses to market shocks across different regimes and directions. Machine-learning shock detection and quantile regression are applied to compare their resilience, co-movement, and conditional behavior under financial stress. Design/Methodology/Approach – An ensemble of unsupervised anomaly detectors (Isolation Forest, Local Outlier Factor, One-Class SVM, Autoencoder) identifies endogenous shocks in the S&P 500 index. Detected shocks are refined through economic and statistical filters to ensure significance and clear event separation. Event-study methods estimate abnormal returns (AR) and cumulative abnormal returns (CAR/CAAR) around shock dates. Quantile regression examines tail-dependent changes in market linkage by interacting asset returns with shock indicators, while controlling for heteroskedasticity across pre-, during-, and post-pandemic regimes. Findings – Gold exhibits consistent downside protection: adverse event-day abnormal returns are limited, CAAR drawdowns remain muted, and partial post-shock recovery patterns are common. In lower quantiles, safe-haven coefficients approach zero or turn negative, indicating weaker co-movement under financial stress. Bitcoin displays pronounced volatility with larger drawdowns during negative shock periods, limited rebound capacity, and intensified market linkage in lower quantiles. Safe-haven features weaken during positive shock events for both assets. Research Implications – Safe-haven behavior is conditional and regime-dependent. Portfolio policy should adopt dynamic allocation strategies keyed to shock intensity. Gold remains a robust defensive anchor, whereas Bitcoin’s role is context-dependent, supporting adaptive diversification strategies.