Pronóstico de la volatilidad usando perceptrones multicapa con funciones adaptativas de activación

Artificial neural networks has been used successfully for forecasting nonlinear time series. In this paper, we present a novel approach to forecast the volatility of financial time series using a multilayer perceptron with adaptive activation functions; model parameters are estimated by maximizing the natural logarithm of the residuals. To guaranty that the variance is always zero or positive, some restrictions are imposed to the neural network. For evaluating the predictive ability of the proposed approach, we compared the forecasts of an ARCH model and the neural network; we found that our approach is able to forecast with more accuracy the volatility that the classical model. Keywords— Non-linear ARCH, Non-linear models, Forecasting models, Heteroscedasticity, Conditional heteroscedastic models, Predictive accuracy.