Performance of Time-Varying Correlation Estimation Methods

This study evaluates and compares alternative time series correlation modeling techniques, using a broad database of 33 variables and 467 asset pairs in nine different asset classes. For each pair of assets a time-varying moving window correlation (MWC) is computed from different moving itional correlation (DCC) time series model, first documenting the closeness of various MWC estimates to DCC, and next evaluating the effectiveness of the models in a portfolio context. We consider four statistical measures of closeness to DCC. According to the concordance correlation coefficient, the Kolmogorov-Smirnov statistic, and the sign agreement test, across all asset pairs under consideration, the shorter to intermediate moving windows (252 days and below) tend to lie closest to DCC; whereas for the mean square error measure, longer windows tend to best match DCC. However, there are some patterns distinct to certain asset classes such as equity and credit, in which both mean square error and concordance correlation coefficient measures of closeness suggest that MWC match DCC estimates at shorter moving windows. In the portfolio management context, the economic closeness test based on 2,802 monthly rebalanced two-asset portfolios shows that generally MWCs are closer to DCC at the longer window lengths