Shipping-Freight Forecasting with Multi-Domain Features: Role of News Sentiment

The shipping market is complex and nonlinear, which makes freight-rate forecasting highly challenging. This study proposes an STE-Informer model that integrates a multisource feature set. First, we collect 120,000 shipping-news articles from several major industry websites and use the FinBERT model to extract sentiment features, capturing market-sentiment fluctuations. Second, we apply technical analysis to construct a set of technical indicators, exploring the intrinsic information embedded in freight indices. Third, we incorporate economic features to reflect the impact of global economic conditions and external shocks on the shipping market. Finally, we integrate sentiment, technical, and economic features into a multidimensional feature matrix and employ the Informer deep neural network for freight-index forecasting. The results show three key findings: (1) short-term sentiment provides the best predictive performance, improving accuracy by 10%–70% compared with medium- and long-term sentiment; (2) the multisource feature set reduces mean squared error by 76.9%, 83.1%, 88.2%, and 92.5% across four shipping markets, respectively, relative to the no-feature baseline Informer(N); (3) the proposed STE-Informer model achieves the best overall performance in shipping-index prediction, ranking first in forecasting the Baltic Dry Index and Baltic Dirty Tanker Index.