Abstract: As an important safe-haven asset, the quantitative description and prediction of the price volatility of gold is of great significance to the risk management decisions for various investors. Based on the standard regression model, principal component analysis, combination forecast, and two model shrinkage methods (i.e., Elastic net and Lasso) are employed to design new volatility prediction models and investigate which method can more effectively use the information of multiple predictors. In addition, the model confidence set (MCS), out-of-sample R2, Direction-of-change (DoC) evaluation methods are used to assess the out-of-sample prediction accuracy of the new models. The empirical results show that no matter which evaluation method is used, compared with other competition models, the out-of-sample prediction accuracy of the two shrinkage models is the best, which can provide a reliable guarantee for the volatility forecast of China's gold futures prices.

Key words: realized volatility; high-frequency data; combination forecast; machine learning; Elastic net; Lasso