Artificial Neural Networks in Space of Stock Returns: Volatility Prediction

Different Generalized Autoregressive Conditional Heteroskedasticity models were estimated over various constituents of S\&P 500 index. Best performing estimated models were then picked up to demonstrate that even in the case when one of these models is dominant, a nonlinear combination of these models may perform better on out of sample data in certain cases. A popular nonlinear modeling technique called Artificial Neural Networks has been used to combine the GARCH volatility predictions along with the multiple metrics such as mean and variance of stock's returns over a specified window. This approach's goal is to be able to balance multiple models that are performing better in different cases such as sudden downturn or relative stability on markets with the metrics about returns to find the most suitable combination. This combination, however, might be highly complex and therefore Artificial Neural Network is estimated to capture this possible relationship. Various similar models have been already studied, that however assumed a constant relationship between GARCH type models predictions. Using window descriptive statistics enables this neural network to change these relationships dynamically in a hope of better improvement. Better volatility prediction is a crucial feature of financial risk management and this approach, despite being simplified in various places of the analysis, might be a contribution to already existing research and is a step toward a different approach to ensemble models. Sticking to a simplified approach of coding ANN so that it could be written almost from scratch is most likely the reason for somewhat unstable results throughout different stocks. However, the uniting factor has been the improved performance in terms of log-likelihood of predicted one step ahead volatilities throughout all analyzed stocks.