COMBINING MULTIVARIATE VOLATILITY FORECASTS IN A HIGH-DIMENSIONAL SETTING: DOES STATISTICAL PERFORMANCE TRANSLATE TO PORTFOLIO PERFORMANCE?

Abstract

Combining multivariate volatility forecasts remains under-researched despite the well-established advantages of forecast combinations for univariate forecasts. Two broad approaches have been proposed – the first constructs combination forecasts using statistical loss functions and the second does so with economic criteria (based on, for example, portfolio selection). However, it is unclear if the advantages of these approaches documented in the literature can be replicated for large asset dimensions and whether statistical performance translates into portfolio performance. Using a cross-section of 100 S&P 500 stocks, combination forecasts from three methods were evaluated against those from individual models and the simple average benchmark based on statistical and portfolio performance. Empirical findings provided support for the use of combination approaches to improve on the statistical performance of individual model forecasts. While portfolios constructed from the combination forecasts generally performed well, there was not necessarily a one-to-one equivalence between statistical performance and portfolio performance.