Forecasting the variance of stock index returns using jumps and cojumps

Abstract

Modeling and forecasting the variance of asset returns is an important issue in many financial applications. Previous studies have examined the roles of both the continuous and jump components of the total variance in forecasting. This paper considers how index-level jumps and cojumps can be used across index constituents for forecasting the variance of index-level returns. A range of jump and cojump detection methods, based on daily and intraday data, are used. Moving beyond the magnitudes of the past index jumps used in existing models, it is found that incorporating the estimated jump intensity from a point process model leads to forecast accuracy gains. Another important contribution is the finding that cojumps across underlying constituent stocks are also useful for forecasting index-level behaviour. Improvements in forecast performance are particularly apparent on the days when jumps or cojumps occur.