Two Are Better Than One: Volatility Forecasting Using Multiplicative Component GARCH Models
A preliminary version of this paper circulated under the title “On the Statistical Properties of Multiplicative GARCH Models” (2016)
51 Pages Posted: 21 Mar 2016 Last revised: 20 Mar 2018
Date Written: March 15, 2018
We examine the forecast performance of multiplicative volatility models that can be decomposed into a short- and a long-term component. First, we show that in multiplicative models, returns have a higher kurtosis and squared returns have a more persistent autocorrelation function than in the nested GARCH model. Second, we provide theoretical and simulation evidence suggesting that the QLIKE loss should be preferred relative to the squared error loss when comparing volatility forecasts. In a Monte-Carlo simulation, we investigate how the multiplicative structure affects forecast performance both in comparison to the nested GARCH model and the popular HAR model. Finally, we consider an application to S&P 500 returns. Based on the QLIKE loss and forecast horizons of two- to three-months ahead, our results show that multiplicative GARCH models incorporating financial and macroeconomic variables improve upon the HAR model.
Keywords: Forecast evaluation, GARCH-MIDAS, Mincer-Zarnowitz regression, volatility persistence, volatility component model, long-term volatility, model confidence set.
JEL Classification: C53, C58, G12
Suggested Citation: Suggested Citation