Dimension reduction techniques are being used in the economic and financial literature as an alternative way to modelling and forecasting the volatility and circumvent the curse of dimensionality. Most of procedures proposed in the literature are developed to extract the main features of returns and assume that the principal/common component for volatilities coincide with the volatility of the principal/common components returns. These procedures are usually outperformed by other methodologies including simpler multivariate volatility models such as EWMA, ORE and RiskMetrics. On the other hand, there are a few proper procedures to extract the main features of the volatility process. However, all of them are extremely sensitive the possible presence of outliers. This project aims to discuss dimension reduction techniques for volatilities to model and forecast the conditional covariance matrix in a large panel of financial time series as well as to propose a proper robust dynamic dimension reduction technique for volatilities. The new procedure will be compared with other procedures proposed in the literature.
News published in Agência FAPESP Newsletter about the scholarship: