SPPformer: A transformer-based model with a sparse attention mechanism for comprehensive and interpretable ship price analysis

Accurate prediction of vessel prices across multiple ship types is crucial for providing scientific decision-making support to shipping enterprises, investment institutions, and policymakers. However, traditional statistical methods and machine learning models face significant limitations in accuracy, generalization, applicability, interpretability, and data coverage, rendering them inadequate for high-precision forecasting in the complex shipping market. Inspired by the successful application of Transformer-based models like ChatGPT across various domains, this study proposes a novel ship price prediction model based on the Transformer architecture—SPPformer. The model integrates architectural optimization, pre-training, and fine-tuning techniques to enable comprehensive price forecasting for multiple ship types. On the data front, this paper consolidated over one million data variables from 12 maritime-related domains to pre-train the model, forming a Basic Model with foundational time-series processing capabilities. Subsequently, domain-specific maritime data were incorporated through fine-tuning to develop the Sectional Model, enhancing its specialization in the shipping sector. For interpretability, the SHAP method was embedded into the SPPformer prediction framework to visualize the impact of feature variables on target variables during the forecasting process. In terms of efficiency, a sparse attention mechanism was introduced by combining Atrous Self-Attention and Local Self-Attention, replacing the global attention mechanism of traditional Transformer, thereby significantly improving training efficiency and solving overfitting issues. The empirical study focused on predicting newbuilding and secondhand vessel prices for dry bulk, container, tanker, and the overall shipping market. The results demonstrate that the SPPformer model outperforms traditional ones in accuracy and interpretability. At the same time, the introduction of sparse attention reduces training time and memory usage by 22.91 % and 26.12 %, respectively, compared to global attention mechanisms. This research provides essential references for shipping enterprises to enhance economic efficiency and for financial institutions to manage risks. It also offers data-driven support and analytical frameworks for governments to regulate market order and promote the stable development of the shipping industry.