Title: Forecast combination view of HAR model
Authors: Andrey Vasnev - University of Sydney (Australia) [presenting]
Adam Clements - Queensland University of Technology (Australia)
Abstract: The heterogeneous autoregressive (HAR) has become a popular model to predict realized volatility. It is simple and delivers good empirical results. Other modifications, e.g., HAR-Q model, were able to improve the results but only marginally. We take a step back and look at the HAR model as a forecast combination model that combines three predictors: previous day realization (or random walk forecast), previous week average, and previous month average. When applying the OLS method to combine the predictors, the HAR model uses optimal weights that are known to be problematic in the forecast combination literature. In fact, the average forecast often outperforms the optimal combination in many empirical applications. We investigate the performance of the individual predictors. We then use equal weights instead of optimal weights and achieve up to 52\% improvement as measured by the mean squared forecasting error. The combination framework opens a door of possibilities to construct the weights to achieve even more significant improvements. Smaller gains are observed in the context of multivariate HAR models when forecasting the covariance matrix of returns. However, these gains are meaningful, as simple combinations approaches avoid the estimation of large dimensional regression models.