Realized volatility forecasting using hybrid neural networks: An application for the Istanbul stock exchange

Volatility forecasting in the financial markets is important in the areas of risk management and asset pricing, among others. In this study, BIST 100's 1-day, 5-day, and 10-day-ahead return volatilities are examined. Two types of hybrid models are utilized to improve individual GARCH-family models' predictions. For the first hybrid model, a group of GARCH-family models is constructed to produce volatility estimates which were then fed into neural network to increase the predictive power. The second hybrid model received GARCH-family models' specifications instead of volatility estimates as inputs for ANN to conduct the learning process. Hybrid neural networks were also fed a set of exogenous, endogenous, and dummy variables. One of the main conclusions is that both hybrid models increased the forecasting precision of individual GARCH-family models while the second hybrid model provided better volatility forecasts for all error measures used in this study. Equal forecast accuracy test also showed that the hybrid models' out-of-sample predictions were significantly better than GARCH-family methods. All model performances deteriorated as forecast horizon was extended, although the steepest decline happened for hybrid models rather than the GARCH-family. Lastly, as the complexity of the neural network architecture was increased, the loss measures for the out-of-sample forecasts improved except on the last case where the network overfitted using the highest number of neurons per hidden layer among the searched hyperparameter grid.