The Emotional Drive of Economic Forecasting: GDP Predictions Based on Large Language Models

Abstract

This paper presents a study on predicting quarterly GDP by integrating the large language model Fin-Bert with a deep learning long short-term memory (LSTM) network, utilizing news articles and public speeches from 2010 to the second quarter of 2024. First, the Fin-Bert model is employed to analyze a dataset comprising 280,139 samples of quarterly news articles and public speeches during this period, calculating sentiment scores for the text. Subsequently, these sentiment scores along with relevant macroeconomic factors are utilized as input features to construct an LSTM prediction model aimed at forecasting real GDP values for the third and fourth quarters of 2024.Through an analysis of the prediction results, this study demonstrates that the model incorporating news text sentiment analysis significantly outperforms alternative approaches in terms of accuracy when predicting quarterly GDP. This result suggests that the use of a large language model to analyze news text sentiment and use it as a base feature input, combined with deep learning techniques can improve the accuracy of quarterly GDP forecasting and provide a new perspective for policy making and economic decision making.