LAUSR

We examine whether it is possible to improve volatility forecasts by simultaneously accounting for parameter instability and model uncertainty. Changes in the regression coefficients and/or the error variance are driven by mixture distributions for state innovations (MIA) of linear Gaussian state-space models. This framework allows us to easily compare models that assume small, frequent as well as models that assume large but rare changes in the aforementioned parameters. To account for model uncertainty, we resort to Bayesian model averaging (BMA), where we average across different specifications selected from a set of predictors that includes a lagged value of realized volatility, financial and macroeconomic data. An empirical application using S&P 500 monthly and quarterly realized volatility data from 1960 to 2014 suggests that averaging over several predictors in a model that allows for breaks in the regression coefficients and the error variance through MIA dynamics results in competitive density and point forecasts. However, compared to a heteroscedastic MIA autoregression without performing BMA, we fail to generate statistically significant more accurate forecasts.