Neural Network Architecture Development for Time Series Forecasting

Author Name(s): Alexander Zakharov, Galim Vakhitov, Zulfira Enikeeva
Author Email:


The paper reveals the results of studies on the development of neural network architecture for time series forecasting. Three types of neural networks were investigated: radial network, convolutional network, and multilayer perceptron. The main objective of this study is to identify, based on a comparative analysis, the architecture of a neural network, which is most suitable for the specifics of forecasting data of a financial nature. In the course of work, from the more than 1000 configurations of neural networks, the most successful ones were selected, with the smallest standard error. The C# programming language and the Visual Studio development environment were chosen in the capacity of the implementation tools. In the work, the back propagation of error algorithm was used as a method for training the neural network. During the training of the network, a dynamic change in the network configuration was carried out. The time series of the financial market were used as the data supplied to the neural network. The work was carried out on the data of exchange rates for 2000-2018. As a result of the study, it was determined that the use of a multilayer perceptron provides higher accuracy. The results of the work can be applied to forecast the values of securities, currency pairs and other financial instruments.


The analysis of time series by machine learning methods has received considerable attention [1], [4], [5]. For this, such methods as Multi-Scale Methods (convolutional neural networks), Time-series Shapelets (Wang, Oates) [2], Time Warping Symbolic Aggregation Approximation (L. Ye and E. Keogh) [3] and radial, convolutional, and multilayer neural networks architectures are used. Also, the research results were highlighted in such works as [6], [7], [8], [9]. Our work can partly be considered as a continuation of [10]; therefore, we will consider data on changes in exchange rates as time series. The main goal of the work is to identify a neural network configuration that is most suitable for forecasting the time series values based on studying the mean square error (MSE) value for a neural network with the architecture developed by us. The work was carried out proceeding from the data on exchange rates for 2000-2018.


This study may be useful in the formation of a methodology for identifying the configuration of neural networks, allowing the forecast for the values of time series with maximum accuracy. As a result of the study, it was determined that the use of a multilayer perceptron provides higher accuracy. The network configurations presented in this paper can be used to forecast exchange rates, stocks and other financial instruments.

270 total views, 1 views today

Download PDF File

About the author: admin