Time series consist of ordered sequences of observations collected at regular temporal intervals, an‎d the ability to forecast these sequences with precision is essential for generating reliable insights an‎d supporting informed strategic decision-making. Among the various forecasting domains, financial time series present particularly formidable challenges because of their pronounced volatility, structural unpredictability, an‎d sensitivity to a wide range of external factors. Although traditional statistical techniques an‎d conventional machine learning algorithms have been widely applied, they often fail to capture the intricate, nonlinear, an‎d highly dynamic behaviors that characterize modern financial markets. In recent years, deep learning approaches, particularly transformer-based architectures, have achieved substantial progress by leveraging self-attention mechanisms capable of modeling long-range temporal dependencies an‎d improving predictive accuracy. The emergence of Large Language Models (LLMs) has introduced additional opportunities for financial time series forecasting; however, since these models were originally developed for natural language processing tasks, their adaptation to the numerical an‎d temporally dependent nature of financial data requires targeted methodological refinements an‎d carefully designed integration strategies. This study examines financial time series forecasting through a systematic eva‎luation of the capabilities an‎d limitations of Large Language Models. For this purpose, cryptocurrency price data are collected, preprocessed, an‎d reformatted to ensure compatibility with LLM-based modeling pipelines. In addition, textual data extracted from cryptocurrency market news are incorporated as a supplementary source of information, enabling an assessment of their influence on short-term price movements. By developing a hybrid methodological framework that integrates specialized pro‎mp‎t-engineering techniques, restructures numerical inputs to enhance model interpretability, an‎d applies multi-step forecasting strategies, this research introduces an approach that differs meaningfully from existing transformer-based an‎d machine learning methods. Through comparative analysis with commonly used time series forecasting techniques, the study establishes a solid foundation for eva‎luating the predictive performance of the proposed framework, clarifying the practical capabilities of Large Language Models in financial applications, an‎d identifying promising directions for their future development.

مهران صفاياني

عليرضا بصيري , حسين فلسفين