An implementation of the LHAR-CJ model with functional coefficients

Abstract

This article aims to compare the forecast performance of the LHAR-CJ model, proposed in Corsi and Renò (2012) and a LHAR-CJ model with functional coefficients for a Vale return series. This new model, instead of estimating fixed coefficients for each variable in the autoregressive model, estimates a functional coefficient that is state dependent, where the state is represented by the lagged realized volatility. In other words, the coefficients are functions of the states of the response variable. We found out that, for this data, the functional coefficients model has a better forecast performance with the right smoothness parameter.