标题：Comparison of Neural Networks and Regression Time Series in Estimating the Development of the Afternoon Price of Palladium on the New York Stock Exchange
摘要：Purpose of the article: Palladium is presently used for producing electronics,industrial products or jewellery,as well as products in the medical field.Its value is raised especially by its unique physical and chemical characteristics.Predicting the value of such a metal is not an easy matter (with regard to the fact that prices may change significantly in time).Methodology/methods: To carry out the analysis,London Fix Price PM data was used,i.e.amounts reported in the afternoon for a period longer than 10 years.To process the data,Statistica software is used.Linear regression is carried out using a whole range of functions,and subsequently regression via neural structures is performed,where several distributional functions are used again.Subsequently,1000 neural networks are generated,out of which 5 proving the best characteristics are chosen.Scientific aim: The aim of the paper is to perform a regression analysis of the development of the palladium price on the New York Stock Exchange using neural structures and linear regression,then to compare the two methods and determine the more suitable one for a possible prediction of the future development of the palladium price on the New York Stock Exchange.Findings: Results are compared on the level of an expert perspective and the evaluator’s – economist’s experience.Within regression time lines,the curve obtained by the least squares methods via negative-exponential smoothing gets closest to Palladium price line development.Out of the neural networks,all 5 chosen networks prove to be the most practically useful.Conclusions: Because it is not possible to predict extraordinary situations and their impact on the palladium price (at most in the short term,but certainly not over a long period of time),simplification and the creation of a relatively simple model is appropriate and the result is useful.
关键词：palladium;artificial neural networks;regression time series;prediction;multilayer perceptron network
