Forecasting Currency Pairs with RBF Neural Network Using Activation Function Based on Generalized Normal Distribution Experimental Results
Identifikátory výsledku
Kód výsledku v IS VaVaI
<a href="https://www.isvavai.cz/riv?ss=detail&h=RIV%2F61989100%3A27510%2F19%3A10242814" target="_blank" >RIV/61989100:27510/19:10242814 - isvavai.cz</a>
Výsledek na webu
<a href="http://apps.webofknowledge.com/InboundService.do;jsessionid=D80209387398B355CBDABC8E214C8966?customersID=Alerting&mode=FullRecord&IsProductCode=Yes&product=WOS&Init=Yes&Func=Frame&DestFail=http%3A%2F%2Fwww.webofknowledge.com&action=retrieve&SrcApp=Alerting&SrcAuth=Alerting&SID=C4yjVMmap6ihI6dRJVN&UT=WOS%3A000510485200002" target="_blank" >http://apps.webofknowledge.com/InboundService.do;jsessionid=D80209387398B355CBDABC8E214C8966?customersID=Alerting&mode=FullRecord&IsProductCode=Yes&product=WOS&Init=Yes&Func=Frame&DestFail=http%3A%2F%2Fwww.webofknowledge.com&action=retrieve&SrcApp=Alerting&SrcAuth=Alerting&SID=C4yjVMmap6ihI6dRJVN&UT=WOS%3A000510485200002</a>
DOI - Digital Object Identifier
—
Alternativní jazyky
Jazyk výsledku
angličtina
Název v původním jazyce
Forecasting Currency Pairs with RBF Neural Network Using Activation Function Based on Generalized Normal Distribution Experimental Results
Popis výsledku v původním jazyce
In this paper, we implement an effective way for forecasting financial time series with nonlinear relationships. We use the artificial neural network of feedforward type for making the decision-process in a company more efficient, more flexible and more accurate. The main objective of this study is to design new method for improving the performance of RBF artificial neural networks. Based on pre-experimental statistical analysis of 1225 financial time series and inspired by GARCH model, RBF neural networks with new shapes of activation functions based on generalized Normal distribution function (GED) are suggested and discussed. Within this study various types of GED activation function in RBF networks are investigated to find the best ones. Firstly, the presence of homoscedasticity and the occurrence of normality of the time series data is investigated. To test our hypothesis about the application of GED distribution in the RBF neural network, we implemented a neural network application (RBFNN) in JAVA. Using the software, we investigate the RMSE error based on the value of p parameter in GED. The optimized size of the p parameter is determined for classic and soft RBF network related to minimal prediction error. We then test our model on 25 financial datasets to explore the contribution of our suggested and implemented method. We also evaluate the forecasting performance of suggested neural network in comparison to established models based on RMSE. Our results show that the proposed approach achieves higher forecasting accuracy on the validation set than available techniques. The suggested modification form of the shape of activation function of the RBF neural network using GED distribution improves the approximation and prediction accuracy of the RBF network models used for financial time series. From performed experiments we find that the optimal size of the parameter p will likely be in the interval (1.4, 2.4) for a standard RBF and less than 2 for the soft RBF.
Název v anglickém jazyce
Forecasting Currency Pairs with RBF Neural Network Using Activation Function Based on Generalized Normal Distribution Experimental Results
Popis výsledku anglicky
In this paper, we implement an effective way for forecasting financial time series with nonlinear relationships. We use the artificial neural network of feedforward type for making the decision-process in a company more efficient, more flexible and more accurate. The main objective of this study is to design new method for improving the performance of RBF artificial neural networks. Based on pre-experimental statistical analysis of 1225 financial time series and inspired by GARCH model, RBF neural networks with new shapes of activation functions based on generalized Normal distribution function (GED) are suggested and discussed. Within this study various types of GED activation function in RBF networks are investigated to find the best ones. Firstly, the presence of homoscedasticity and the occurrence of normality of the time series data is investigated. To test our hypothesis about the application of GED distribution in the RBF neural network, we implemented a neural network application (RBFNN) in JAVA. Using the software, we investigate the RMSE error based on the value of p parameter in GED. The optimized size of the p parameter is determined for classic and soft RBF network related to minimal prediction error. We then test our model on 25 financial datasets to explore the contribution of our suggested and implemented method. We also evaluate the forecasting performance of suggested neural network in comparison to established models based on RMSE. Our results show that the proposed approach achieves higher forecasting accuracy on the validation set than available techniques. The suggested modification form of the shape of activation function of the RBF neural network using GED distribution improves the approximation and prediction accuracy of the RBF network models used for financial time series. From performed experiments we find that the optimal size of the parameter p will likely be in the interval (1.4, 2.4) for a standard RBF and less than 2 for the soft RBF.
Klasifikace
Druh
J<sub>SC</sub> - Článek v periodiku v databázi SCOPUS
CEP obor
—
OECD FORD obor
10200 - Computer and information sciences
Návaznosti výsledku
Projekt
—
Návaznosti
S - Specificky vyzkum na vysokych skolach
Ostatní
Rok uplatnění
2019
Kód důvěrnosti údajů
S - Úplné a pravdivé údaje o projektu nepodléhají ochraně podle zvláštních právních předpisů
Údaje specifické pro druh výsledku
Název periodika
Journal of Multiple-Valued Logic and Soft Computing
ISSN
1542-3980
e-ISSN
—
Svazek periodika
6
Číslo periodika v rámci svazku
33
Stát vydavatele periodika
US - Spojené státy americké
Počet stran výsledku
24
Strana od-do
539-563
Kód UT WoS článku
000510485200002
EID výsledku v databázi Scopus
2-s2.0-85078211124