Application of nonlinear dynamics and machine learning methods for forecasting economic volatile processes

The emergence of new computer technologies has made it possible to study (visualize) virtually any complex phenomena and processes literally on the display screen. The development of economic and mathematical modeling is influenced by the latest mathematical methods of nonlinear dynamics applied to any field and subject of research. The methods of classical statistics for forecasting economic time series are based on the mathematical apparatus of econometrics. This basing is carried out under the assumption that the observations making up the predicted time series are independent, due to which the necessary subordination to the normal law is satisfied. The latter, however, is the exception rather than the rule for financial and economic time series that have so-called long-term memory. The problem of forecasting and the closely related problem of assessing future economic risks in the conditions of turbulence in the development of the Russian economy, which has manifested itself in recent years, have become especially acute. In the conditions of observed turbulence, economic dynamics become poorly predictable by traditional methods and nonlinear. The direction (growth or decline) of indicators often changes. The research presented in the paper was carried considering the fact that by now there are no complete theories of forecasting time series with memory, which determines the relevance and need to develop new mathematical methods and algorithms to identify the possible predictability of time series with memory and build adequate predictive models. All of the above points to the relevance of developing a qualitatively new methodological approach that ensures the formation of informed management decisions in conditions of uncertainty and risk. The object of the study is time series of economic indicators. The subject of the study is the mathematical, statistical and instrumental apparatus of systems to support management decision-making and forecasting in the economy. The article describes the procedure for constructing models and presents the results of their testing for a linear cellular automaton, the exponential smoothing method, and the Holt–Winters method. During the study, large volumes of data were processed. At the same time, daily, weekly and seasonal indicators are considered. This made it possible to adequately describe and predict the nonlinear dynamics of time series. Using all three levels of analysis simultaneously allows you to get a more complete understanding of the dynamics of time series indicators. In the article, the proposed tools are tested using the example of platinum prices. The developed economic and mathematical apparatus is also applicable to other time series characterizing certain economic variables.