Prediction of export sales volumes using machine learning in conditions of unstable demand

Abstract

The relevance of the study stems from the increasing instability of demand in foreign markets, the volatility of macroeconomic and market factors, the disruption of logistics chains, and the growing uncertainty surrounding conditions for implementing export activities. In such a situation, traditional approaches to forecasting export sales volumes do not provide sufficient accuracy and adaptability, which makes the application of machine learning methods relevant, as they can account for the multifactoriality and nonlinearity of market processes. The purpose of the article is to scientifically substantiate and develop approaches to forecasting export sales volumes based on machine learning methods in conditions of unstable demand in order to increase the accuracy of forecast estimates and the quality of management decisions in the field of export activities. Methods. The study uses systemic and analytical approaches, methods of theoretical generalization, structural-logical analysis and comparison to systematize modern approaches to forecasting economic indicators. A conceptual analysis of forecasting models is used, taking into account multifactoriality, nonlinearity and instability of time series. Results. The impact of macroeconomic, market, political, regulatory, and logistical factors on demand instability and the dynamics of export sales is analyzed. Methodological approaches to the application of regression, ensemble, and neural network machine learning models for forecasting in the presence of nonlinearity and structural changes are summarized. Problems with the practical use of such models are identified, including data instability, structural breaks in time series, and limited interpretability of results. Conclusions. The feasibility of using machine learning methods for forecasting export sales volumes under unstable demand is confirmed as a tool for improving the analytical quality of forecasts and the effectiveness of management decisions. The need to integrate such models into the export management support system of enterprises is substantiated. Prospects for further research include the development of interpreted and hybrid forecasting models, the improvement of methods for analyzing unstable time series, and the formalization of procedures for automatic detection of structural breaks in the dynamics of export sales.