Abstract:
This study focuses on improving forecasting models for Ukraine's macroeconomic indicators
under conditions of volatility and exogenous shocks. The scientific novelty lies in the integration
of neural network theory with the principles of dynamic system identification. The LSTM
architecture effectively models significant time lags and resolves the vanishing gradient problem.
Model parameters are determined using iterative optimization algorithms to minimize the loss
function. Computational experiments on statistical data confirm the superiority of deep learning
methods over linear approaches. The results justify the use of non-linear models to account for
economic inertia and enhance the accuracy of long-term forecasting
