Forecasting economic indicators with LSTM neural networks and graph-based correlation models

This paper explores the application of Long Short-Term Memory (LSTM) neural networks for forecasting Ukraine’s macroeconomic indicators, particularly in the context of structural changes caused by wartime disruptions. The limitations of traditional methods, such as Autoregressive Integrated Moving Average (ARIMA) and Vector Autoregression (VAR) models, which struggle to account for nonlinear dynamics and adapt to abrupt changes, are analyzed. The study substantiates the feasibility of using LSTM as a more flexible approach capable of capturing complex temporal dependencies and improving forecasting accuracy. Graph-based correlation models were employed to analyze interdependencies between macroeconomic indicators, allowing the identification of key economic clusters and the most influential variables. Experimental testing of the model was conducted on data on the economic state of Ukraine, and the forecasting results were compared with the baseline models (naive forecast and ARIMA). Accuracy evaluation demonstrated that LSTM outperforms traditional approaches in terms of Mean Absolute Error and Root Mean Squared Error, especially under conditions of economic instability. Analysis of the results in the period after the start of the full-scale war revealed difficulties associated with changing economic relations and disruption of previous trends, which reduced the accuracy of forecasts. Ways of adapting the model are proposed, in particular, the introduction of regime variables reflecting the influence of wartime factors and external financial factors, the application of a step-by-step learning mechanism, as well as the use of correlation graphs to improve the selection of input variables. The results confirm the effectiveness of LSTM-based macroeconomic forecasting and demonstrate that graph-based correlation models can improve model adaptability in periods of economic uncertainty. The proposed methods may contribute to further advancements in forecasting models, particularly in accounting for crisis scenarios and structural shifts in the economy.