Title: Large scale portfolios under transaction costs and model uncertainty: Mixing of high and low frequency information
Authors: Nikolaus Hautsch - University of Vienna (Austria)
Stefan Voigt - WU (Vienna University of Economics and Business) (Austria) [presenting]
Abstract: A Bayesian sequential learning framework is proposed for high-dimensional asset allocations under model ambiguity and parameter uncertainty. We consider portfolio allocations maximizing predictive expected utility after transaction costs, optimally balancing implementation shortfall and adjustments due to updated information. The unifying framework allows for time-varying mixtures of predictive return distributions which may exhibit fat tails, resulting from high- and low-frequency data. The model is estimated via MCMC methods and allows for a wide range of data sources as inputs. We consider predictive models resulting from high-dimensional Wishart approaches for high-frequency based blocked realized kernels, low-frequency based multivariate stochastic volatility factor models and regularized daily covariance estimates. Employing the proposed framework on a large set of NASDAQ-listed stocks,we observe that time-varying mixtures of high- and low-frequency based return predictions significantly improve the out-of-sample portfolio performance compared to individual models and outperform the naive 1/N-allocation in terms of Sharpe ratio and utility-based measures. Bootstrapping the optimization procedure shows that our results are robust with respect to the choice of the asset universe. We show that regularization of turnover is crucial in large dimensions and illustrate that the relative contribution of high-frequency data and low-frequency data strongly varies over time.