The CBOE volatility index (VIX) is a representative barometer of the overall sentiment and volatility of the financial market. This paper seeks to apply random forest and its variable importance measure to forecasting the VIX index. Compared to the previous literature which has found it difficult to outperform the pure HAR process in terms of forecasting the VIX index due to its persistent nature, random forest can produce forecasts that are significantly more accurate than the HAR and augmented HAR models for multi- days forecasting horizons. This paper shows that the forecasting accuracy of random forest could be further improved by systematically selecting the optimal number of the most important covariates from a dataset of 298 macro-finance variables, while using the Boruta algorithm which ranks the variables based on random forest’s variable importance measure. The superior predictability of this method is more evident with longer forecasting horizons.