DCC-GARCH for Volatility and Correlation Forecasting with Exogenous Market Signals

Sammanfattning

Financial markets exhibit time-varying volatility and cross-asset correlation. To construct and optimize a risk-adjusted portfolio and account for underlying collinearity between se- curities, a good understanding of these two dynamics is important. In this study global equity indices are modeled with the purpose of improving the forecasting accuracy of time-varying correlations, as well as distinguishing the difference between global multi- variate and pairwise settings. A core part of this is the difference between a DCC-GARCH baseline model and its numerous exogenous market signal customizations. Examples of these are the volatility index, dark-pool activity index and credit spreads. Smaller models are compared with the full model to question the efficacy of the multivariate approach. The results show that the inclusion of some exogenous variables provide marginal im- provement in predictive accuracy if added right, whilst some worsen it, either numeri- cally or through added complexity. Credit spreads show modest but mostly beneficial improvements, both for in sample fit and out-of-sample forecast errors for both the full multivariate model and for the pairwise models. The volatility index showed nuanced results, improving correlation forecasts for markets pairs such as US-Europe, while wors- ening predictability for US-Asia pairs. DIX shows the weakest and least consistent effects, suggesting dark pool activity carries limited information for global correlation dynamics in our model. A recurring finding is that the DCC-GARCH baseline already captures much of the dynamics that the exogenous signals might otherwise explain. Furthermore, the model dimensionality affects the magnitude of market signal changes. A recurring theme across all signals is that the extent of improvement is small.