Semiparametric Approaches to the Prediction of Conditional Correlation Matrices in Finance

We consider the problem of ex-ante forecasting conditional correlation patterns using ultra high frequency data. Flexible semiparametric predictors referring to the class of dynamic panel and dynamic factor models are adopted for daily forecasts. The parsimonious set up of our approach allows to forecast correlations exploiting both estimated realized correlation matrices and exogenous factors. The Fisher-z transformation guarantees robustness of correlation estimators under elliptically constrained departures from normality. For the purpose of performance comparison we contrast our methodology with prominent parametric and nonparametric alternatives to correlation modeling. Based on economic performance criteria, we distinguish dynamic factor models as having the highest predictive content.