- The paper introduces a GARCH-DCC-Copula model combined with Markov switching to capture extreme events and regime shifts.
- It employs ARMA, GARCH, and copula functions to address volatility clustering, tail dependence, and non-normal asset returns.
- The findings demonstrate improved portfolio performance, risk prediction, and factor weighting in dynamic allocation strategies.
Dynamic Asset Allocation with Extremes, Tail Dependence, and Regime Shifts
This research introduces a comprehensive framework for dynamic asset allocation, incorporating extreme events, tail dependence, and regime shifts to enhance risk management and portfolio construction. The methodology employs a GARCH-DCC-Copula model to capture non-normal asset return characteristics and a Markov switching model to predict real-time risk regimes.
Modeling Non-Normal Asset Returns
The study begins by acknowledging the limitations of traditional portfolio construction techniques that assume multivariate normality. To address this, the authors propose a GARCH-DCC-Copula model to account for several stylized patterns observed in asset returns.
The authors construct a representative static asset allocation (SAA) comprising 50% equities (MSCI World), 40% fixed income (Deutsche Bank Aggregate Bond index), and 10% alternatives (S&P/GSCI) to proxy the global capital market. They use ARMA models to address return serial correlation, GARCH models to deal with volatility clustering and extreme outliers, and DCC to capture time-varying dynamic correlations. Copula functions are used to account for tail dependence.
The paper emphasizes that asset returns exhibit serial correlation at the asset class level, volatility clustering, and heavier tails than predicted by a normal distribution. It uses ACF and PACF to determine the ARMA structure of asset returns. The GARCH(1,1) model is employed to capture volatility clustering, with the GJR-GARCH extension considered for asymmetric risk modeling. However, due to statistically insignificant leverage coefficients for two out of three assets, this feature was excluded from the final model.
Dynamic conditional correlation (DCC) models are used to address the limitations of constant conditional correlation (CCC) models. The paper demonstrates that CCC can be misleading as correlations decrease over time, while DCC provides a more balanced approach. The weighted average portfolio correlation (WPC) calculation is also discussed, highlighting the non-linear relationship between asset volatility and pairwise correlation.
Copula models are used to capture tail dependence, showing that tail dependence coefficients are higher than Pearson correlation coefficients. The t-Copula is preferred for its ability to capture dependent extreme values.
Predicting Global Financial Market Risk
The research uses a GARCH-DCC-Copula model to predict the risk of the global capital market. The model incorporates ARMA(1,1) for conditional mean, GARCH(1,1) for conditional variance, DCC for dynamic conditional correlation, and t-Copula for multivariate tail dependence.
The GARCH-DCC-Copula model is fitted using an expanding window of daily returns, re-estimated monthly. The estimated parameters are used to simulate future returns and calculate risk metrics such as CVaR and volatility. The results show that the model effectively captures the downside risk, with CVaR and volatility peaking during the 2008 financial crisis. The model also provides insights into implied correlations between asset classes.
The accuracy of the risk and correlation predictions is assessed by comparing them with look-ahead estimates. The GARCH-Copula model shows higher predictive power than a naive model for all three asset classes, particularly for commodities. The GARCH-DCC-Copula model also dominates GARCH-CCC-Copula and naive estimates in predicting correlation.
GARCH-DCC-Copula Model for Asset Allocation
The GARCH-DCC-Copula model is extended to a real-life 11-asset class application to assess its ability to build better risk models and more efficient portfolios. The investment universe includes US large-cap equity, US small-cap equity, international equity, emerging markets equity, REITs, US treasuries, US high yield bonds, investment-grade sovereign bonds, EM credit, commodities, and gold.
Various portfolio construction techniques are compared, including naive diversification strategies (equally weighted, inverse volatility, risk parity) and sophisticated diversification strategies (maximum diversification, minimum tail dependence). Risk minimization strategies, such as global minimum variance, minimum variance-tail dependence, and minimum CVaR, are also evaluated.
The results show that portfolios constructed using the GARCH-DCC-Copula model consistently produce higher ex-post returns and reduce both variance-based risk and tail risk. The model also improves portfolio Sharpe ratios, reduces downside risk, and increases diversification.
A case study on rising interest rates demonstrates how the model can better capture fast-moving economic environments and their implications for asset allocation. The GARCH-DCC-Copula model correctly predicted heightened downside risk for bonds, leading to an underweighting of bonds compared to traditional risk models.
Global Tactical Asset Allocation (GTAA) Strategies
The research extends the analysis to GTAA strategies, incorporating return prediction models. Three sets of return prediction models are used: naive alpha (short-term average), naive alpha (long-term average), and a GTAA model using variance risk premium (VRP).
The VRP, defined as the difference between market-implied risk and realized risk, is used to forecast asset returns. A simple forecasting model based on linear regression is constructed, and the information coefficient (IC) is used to measure the performance of the GTAA model.
The paper also discusses portfolio construction techniques for alpha strategies, including mean-variance optimization (MVO) and mean-CVaR optimization (MCVaR). Scenario-based optimization via fractional programming is also mentioned as a tool for solving the MaxSharpe or MaxReturnCVaR problems.
Predicting Market Risk Regime
The study employs a dynamic Markov switching model to predict whether the market is in a high-risk or low-risk regime in real-time. The model incorporates VRP as a predictor of risk regimes and allows for time-varying transition probabilities and regime heteroskedasticity.
Four kinds of regime probabilities are defined: one-step-ahead prediction, filtered, smoothed, and true out-of-sample. The paper emphasizes the importance of using true out-of-sample regime probabilities to avoid look-ahead bias.
The results show that the real-time global financial risk regime indicator has considerable predictive power for asset returns. The paper finds strong evidence that asset classes behave differently in different regimes, which has significant implications for asset allocation decisions. The combination of risk and correlation regimes creates a four-state model, which provides further insights into asset allocation strategies.
The research demonstrates how the predicted real-time risk regimes can enhance the GTAA model, improving the average performance (IC) and risk-adjusted performance.
Risk Premia Allocation and Factor Weighting
The GARCH-Copula-Regime Switching risk model is applied to five simple quantitative stock selection factors (value, momentum, quality, size, and low volatility) to demonstrate its real out-of-sample applicability.
The performance of these factors is analyzed under different risk and correlation regimes. The results show that value and quality tend to perform strongly in high-risk states, while momentum and size perform better in low-risk states. The low-volatility factor, however, tumbles in a high-risk environment.
The paper also demonstrates how regime prediction can improve factor return prediction, using a style rotation model. The model incorporates VRP and the pure out-of-sample predicted high-risk/low-risk regime indicator, improving the average performance (IC) and risk-adjusted performance.
Finally, the research combines the style factor return prediction model, the GARCH-DCC-Copula risk model, and maximum mean-CVaR optimization to construct an active and dynamic factor rotation model. The results show that the factor rotation strategy exhibits higher return and risk compared to risk-based strategies, delivering better performance than risk parity and maximum diversification.
Conclusion
This paper presents a comprehensive framework for dynamic asset allocation and factor weighting, incorporating extreme events, tail dependence, and regime shifts. The GARCH-DCC-Copula model and the Markov switching model provide valuable tools for predicting risk and correlation regimes, improving portfolio construction and enhancing return prediction ability. The findings highlight the importance of accounting for non-normal asset return characteristics and adapting investment strategies to changing market conditions.