Dynamic Neural Network for Business and Market Analysis

Source of Publication

Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)


© 2019, Springer Nature Switzerland AG. The problem of predicting nonlinear and nonstationary signals is complex since the physical law that controls them is unknown and it is complicated to be considered. In these cases, it is necessary to devise nonlinear models that imitate or learn the rules of behavior of the problem and can be developed based on historical data. For this reason, neural networks are useful tools to deal with this type of problem due to their nonlinearly and their capacity of generalizing. This paper aims at exploring various types of neural network architectures and study their performance with time series predictions. Predictions on two sets of data (of a very different nature) will be made using three neural networks including multilayer perceptrons, recurrent neural network and long-short term memory varying some important parameters: input neurons, epochs and the anticipation with which the predictions are made. Then, all results will be compared using standard metrics. As a conclusion, the influence of the type of series under study is more important than the parameters considered in what concerns the performance. The management of the memory in the networks is a key to its success in the prediction of S&P 500 and electrical power time series.

Document Type

Conference Proceeding



First Page


Last Page


Publication Date