A Comparative Study of Time-Series Models Used in Artificial Intelligence Forecasting Approaches

Forecasting dynamics in financial markets remains a central yet unresolved challenge due to their inherent volatility, nonlinear dependence, and the presence of structural breaks and seasonality. The accurate modeling of stock price movements is not only of theoretical significance but also of considerable practical relevance for investment strategy design, portfolio optimization and systemic risk management. In this study, we investigated a comprehensive comparative analysis of forecasting techniques, encompassing both classical statistical models and modern machine learning approaches specifically adapted for time series prediction. Traditional econometric methods, as generalized autoregressive conditional heteroskedasticity (GARCH) and autoregressive integrated moving average (ARIMA) were systematically evaluated alongside conventional machine learning algorithms, including linear regression and support vector machines (SVM). Beyond these baselines, we assessed the predictive capacity of advanced neural network architectures, with particular emphasis on long short-term memory (LSTM) networks and convolutional neural networks (CNN), which are designed to capture long-range temporal dependencies and nonlinear feature interactions. Empirical experiments conducted on real stock market datasets allow for a rigorous performance assessment under diverse market regimes. The results demonstrated differentiated strengths across methods, where statistical models retain interpretability and robustness, while deep learning approaches yield superior accuracy in highly volatile environments. The study concludes with evidence-based recommendations concerning methodological suitability for varying forecasting horizons and financial application scenarios.