Combining neural network and HAR forecasts of perceived volatility

This paper examines a combination of HAR and neural networks methods to better predict perceived volatility and, consequently, to more efficiently manage risk. To carry out the projections, combinations and tests, the series of perceived volatility of Ibovespa was collected between 2000 and 2018, producing a sample of 4,530 observations. The main results show that the combination of both models better predict perceived volatility, which can be interpreted as an efficiency gain for risk management. In addition, this article also evaluates the performance of the models, considering the profitability of trading with options. For the case of profitability, combinations of linear and nonlinear models present better performance.