Possibility of using large language models in time series prediction

This thesis investigates how time series prediction can be improved by using large language models. Discrete time series with EUR/USD workday price is chosen, furthermore two article sources whose influence on EUR/USD price is quantified are chosen, fxnewssite representing specialized news and bbc representing general news. Influence prefilter is constructed allowing chosing relevant articles from general news. Two times (without and with influence prefilter) 444 articles from each source (fxnewssite and bbc) in period between 01. 10. 2021 and 31. 12. 2023. are analyzed. Expected influence of articles on EUR/USD price is quantified by large language model gpt-3.5-turbo-1106 with four created templates. Two machine learning models are trained for time series and selected combination of templates, base model only with prices and endogenous variables created by feature engineering from dates and prices and enhanced model extended by attributes created using large language model. The difference in prediction is measured. In the cases analysed, where relevant and high-quality textual source fxnewssite was available, an improvement of five day forecast was achieved ranging from 12.38 to 15.35 percent when using influences calculated always by one template without using the influence prefilter. For bbc source the influence prefilter causes shift of average prediction from individual templates from improving by 0.2675 percent to improving by 2.0125 percent.