Borsada Optimum Yapay Sinir Ağının Belirlenmesi

Abstract

Artificial intelligence applications are widely used in the field of finance as in all fields. In the field of finance, artificial neural networks are the most commonly used artificial intelligence methods for the analysis of time series. The fact that artificial intelligence methods make similar inferences to the human brain or that they are systems based on learning is the most important of their contributions to analysis. Of course, the development process of these methods is closely related to the development of technology should not be forgotten. In other words, despite all the advantages, the demand for the use of methods is now due to the fact that transactions can be done more quickly and simply. Past studies of artificial neural networks show that many of them make successful predictions. However, it is also known that the results are different. This is because there are elements that can affect the performance of the system, such as network architectures, number of layers, and educational algorithms. In this study, the optimum artificial neural network that should be established in Istanbul Stock Exchange is emphasized. How a network structure will work with better performance in Borsa İstanbul is examined. For this purpose, many analyses were carried out with network architectures, hidden layer numbers and educational algorithms. In the analysis, fundamental and technical data were used as input variables in order to predict the BIST100 index. The main variables are US Dollars, interest rates on deposits and money supply (M2). The technical variables are; MACD, RSI, Momentum and Stochastic. As a result of the study, it was observed that the use of feed-forward networks and Bayesian Regulation training algorithm would be appropriate in the analyzes made with BIST100 index. In addition, it was observed that increasing the number of hidden layers increases the performance, but it is observed that performance increase slows down if more than 5 layers are selected. Therefore, it is considered that the ideal number of layers should be 5 against the speed of the system and the problems of memorization of the network.