Dry Bulk Freight Rate Prediction and Interpretive Analysis: Transformer-Based Hybrid Model with Domain Adaptation

Shipping is vital to global trade but the dry bulk market faces high volatility because of complex factors like economic conditions and geopolitics. Traditional models struggle to capture nonlinear rate fluctuations and offer interpretability. To overcome this, our study integrates shipping, financial, and commodity data. Using the Maximal Information Coefficient (MIC) algorithm, 4,281 time series (over 4.53 million points) were selected to pre-train a general maritime time series model (MTS) based on the Transformer. A fine-tuning dataset of 248 high-MIC series (>0.2) and five binary event indicators (over 0.073 million points) was then used for domain adaptation, creating DA-MTS. To better model short-term changes, a long short-term memory (LSTM) network with Grey Wolf Optimization was applied to DA-MTS residuals, yielding a hybrid DA-MTS+LSTM model. Empirical results show DA-MTS+LSTM achieves high accuracy in forecasting the Baltic Dry Index (MAE = 0.16, RMSE = 0.21, MAPE = 5.26%), the Baltic Capesize Index (MAE = 0.21, RMSE = 0.27, MAPE = 6.67%), and the Baltic Panamax Index (MAE = 0.16, RMSE = 0.20, MAPE = 5.12%). The study also analyzes the short- and long-term impacts of exogenous variables and sudden events. This work enhances prediction accuracy and stability while extending Transformer applications to time series analysis, offering robust theoretical and practical decision-making support.