Volatility forecasts by clustering: Applications for VaR estimation

It is well known that volatility has time-varying and clustering characteristics. The information content of volatility clustering is particularly important in turbulent periods, such as the stage of financial crisis. How to fully mine the implicit information within clusters to predict the volatility in the future is a rarely discussed issue. In this paper, we put forward a partition model to segment volatility into non-overlapping clusters by Fisher's optimal dissection methodology. Using this model, we can quickly identify the points of structural changes in volatility. By utilizing the information of the nearest cluster, we can perform point estimation and interval estimation on future volatility. In the end, we conduct some empirical examples based on the returns of S&P 500, DAX 30 and FTSE 100 index. We find that our method can improve the volatility forecast and VaR estimations.