Internal Research Seminars

We consider the problem of identifying the optimal subset of past observations to be used for estimating the covariance matrix of asset returns when computing the global minimum variance portfolio. While the usual approach simply employs either all the available past returns or a rolling window of fixed size, we argue that such solutions are inefficient, and we instead propose to dynamically adjust the estimation window. To this end, we employ the PELT changepoints detection algorithm to identify outdated observations. We then set up a cross-validation procedure to determine the minimum number of upto-date observations that individual assets must have to be considered when estimating the covariance matrix. Assets that do not reach the selected threshold are excluded from the optimization and assigned equal weights. In an empirical application, our solution achieves a higher Sharpe ratio compared to both an expanding and rolling window estimation of the global minimum variance portfolio, and it also compares well with other benchmarks.