Improved Architecture and Configurations of Feedforward Neural Networks to Increase Forecasting Accuracy for Moments of Finite Normal Mixtures

Abstract

The methodology of forecasting using neural networks for modified data with original values replaced by discrete ones was successfully applied in previous authors’ works for mixture moments approximated with method of moving separation of mixtures. Previously the uniform architecture has been used for all analyzed time-series. This paper proposes an improved type of neural network architecture with grid-search hyper-parameter tuning for expectation, variance, skewness and kurtosis. It allows increasing the value of prediction accuracy for some combinations of forecast periods and number of gaps up to 99.7%. Numerous tables and plots are presented to better demonstrate the results.