Application of recurrent neural networks to forecasting the moments of finite normal mixtures

The article compares the application of feedforward and recurrent neural networks to forecasting continuous values of expectation, variance, skewness, and kurtosis of finite normal mixtures. Fourteen various architectures of neural networks are considered. To increase training speed, the high-performance computing cluster is used. It is demonstrated that the best forecasting results based on standard metrics (root-mean-square error, mean absolute errors, and loss function) are achieved on the two LSTM (Long-Short Term Memory) networks: with 100 neurons in one hidden layer and 50 neurons in each three hidden layers.