Estimating Stock Index Volatility in India

Abstract

Since we find that GARCH(1,1) model fails to provide accurate one day volatility forecasts, we attempt to provide better forecasts using autoregressive neural networks. We hypothesize that due to their nonlinear nature, autoregressive neural networks are more effective in capturing the distribution of realized volatility. We use radial-basis-function and feed-forward back-propagation autoregressive neural networks to make one day volatility forecasts, and find that an overwhelming majority of simulated networks produce forecasts that are in order of magnitude more accurate than GARCH(1,1) model is. In addition, we learn that in India, realized volatility does not have an especially long memory, and approximately one week of trading sessions holds all the information a researcher may require to generate one day volatility forecasts.