Comparative Studies of Self-organizing Algorithms for Forecasting Economic Parameters

Volodymyr Lytvynenko, Department of Informatics and Computer Science, Kherson National Technical University, Kherson, Ukraine, Olena Kryvoruchko, Irina Lurie, Nataliia Savina, Oleksandr Naumov, Mariia Voronenko
2020 International Journal of Modern Education and Computer Science  
This manuscript presents the economic research results based on their input-output characteristics and functional description with inductive modeling methods and tools. There are a wide plethora of methods to be used for solving this type of problem, including various neural network models, linear and nonlinear regressions, reference vectors' methods, fuzzy models, etc. The main disadvantage of these methods is that the obtained models cannot always interpret and obtain a model of optimal
more » ... el of optimal complexity. Unlike the mentioned methods and tools, the group method of data handling (GMDH) allows building models directly from a data sample without the attraction of additional a priori information. This algorithm admits finding internal dependencies in the data and determining optimal model complexity. There is a broad range of iterative GMDH algorithms that have been developed and studied. Oversampling algorithms are applicable for solving the structural identification problems for a limited number of arguments. Iteration algorithms are suitable for solving tasks with many arguments, but they do not guarantee proper structure development. Multi-row GMDH iteration algorithms are the most popular ones. However, they have several sufficient defects, such as informative argument loss or non-informative argument inclusion, as well as a polynomial degree of exponential growth. In this context, the applicability of the GMDH-based iterative and combined architectures for solving the model's interrelation problems between a volume of capital investments and GDP by activity types in the transport branch is considered. The determination coefficient is utilized for the estimation of the obtained models based on a complicated evaluation procedure. The Kolmogorov-Smirnov criterion estimates the model's adequacy. The F-criterion Fisher assesses the significance of polynomial models. The demonstrated results proved that the combined iterative and combinatorial algorithms turned out to be the most effective solution for all evaluation criteria.
doi:10.5815/ijmecs.2020.06.01 fatcat:v7xrjdv3dze7xpfqzt4jqn3vre