A data-filtered sentiment analysis model for economic forecasting

Abstract

In behavioral economics, sentiments influence decision-making processes, with positive sentiments tending to underestimate risks and negative sentiments overestimate them. At the individual level, these sentiments shape economic behavior in ways that collectively influence broader economic dynamics. With the proliferation of textual data and advancements in natural language processing techniques, the analysis of economic sentiments has garnered growing attention, with large language models showing superior analytical capabilities. However, unlike many machine learning tasks where true labels are available through human annotation, sentiment analysis encounters challenges in obtaining true labels due to the psychological biases and inconsistencies inherent in human assessments. To address this issue, this study introduced a data filtering methodology to enhance data reliability and developed a robust sentiment analysis model tailored to the Japanese economy. The findings revealed that our model not only outperforms existing models in terms of generalization capability across diverse datasets — achieving RMSE values of 0.09–0.11 and classification accuracies of 0.83–0.88 — but also effectively captures fluctuations in other quantitative economic indicators, as evidenced by Euclidean distances of up to 1.56, which is smaller than the records 4.33 and 4.24 of the existing models. Moreover, a statistically significant correlation between qualitative and quantitative economic indicators was identified, highlighting the potential of qualitative indicators in predicting economic conditions.