Forecasting moments of finite normal mixtures using feedforward neural networks

Modeling and analysis of nonstationary data flows in real systems of various types can be effectively performed using finite local-scale normal mixtures. Approbation of the prediction methodology developed by the authors is carried out on the example of time-varied moments of the mixed probability model. Within this approach, values of the initial continuous time-series are replaced with the discrete ones and then modified samples are analyzed with a neural network. For short-term forecasting, the accuracy of more than $80\%$ is demonstrated. Feedforward neural network is implemented using the Keras deep learning library, the TensorFlow framework, and the Python programming language.