공분산행렬 추정방법과 공매도제한 제약의 자산배분성과에 대한 영향분석

This article tackles how to forecast the covariance matrix of returns. We compare theperformance of different methods of forecasting covariances, with an eye to judging whichmodels improve the ability to optimize portfolio risk. We tested the forecasting performanceof the sample covariance matrix, the constance correlation model (CCOR), the constantcovariance model (CCOV) and the estimation method using random matrix theory (RMT). We find that the forecast performance did not show significant difference among theestimation methods. But they had different performances in the context of optimizedportfolios. We conduct two types of experiments, the global minimum variance portfoliounder each model and minimizing tracking error portfolio. We find that minimum varianceportfolios based on CCOR and RMT method have annualized standard deviations 19.15%and 17.87%, respectively, which are lower than 22.97% when using the sample covariancematrix. The better risk-minimization performance using CCOV and RMT method holdsagain in the experiment of minimizing tracking error volatility. In addition, we tested the effect of short-sale constraint on the portfolio performance. The short-sale constraint typically shrinks the larger elements of the covariance matrixtowards zero, which leads to a more precise estimate. But the constraint introducesspecification error. The net effect depends on the trade-off between sampling and specificationerrors. We demonstrate that with the constraint in place, minimum variance portfoliosconstructed using the sample covariance matrix perform as well as those constructed usingcovariance matrices estimated using CCOV and RMT methods.