Find in Library

Forecast combination models have been broadly studied and often used to improve forecast accuracy. This article presents a new non-linear composite model to forecast the volatility of asset returns. Our model is composed of a set of GARCH models fitted to a time series dataset using different loss functions, with the aim of capturing different features of volatility dynamics. Individual forecasts are combined by using either the simple arithmetical average method or an artificial neural network. The proposed model is used to forecast the monthly excess returns of S&P500 time series, finding that this new approach is able to forecast volatility with more accuracy than each individual GARCH model considered.