A comparative study on shipping accident damage compensation prediction based on SVR and a BP neural network

To address the challenges of institutional gaps in liability regimes, small sample data, and nonlinear regression in shipping accident damage compensation assessment, this study proposes SVR and BP neural network models and compares their performance in small sample scenarios. Using data from 17 typical shipping accidents (1991—2020) and seven key risk factors (e.g. leakage volume and gross tonnage), the models were built and optimized. The results indicate that SVR outperforms BP neural networks when evaluated using the root mean squared error (RMSE) and mean absolute error (MAE) metrics. SVR's superior generalizability via structural risk minimization and kernel mapping makes it more effective for small sample nonlinear regression. Conversely, BP neural networks suffer from local optima trapping and parameter sensitivity, leading to lower prediction accuracy. This study offers a data-driven solution for rapid shipping accident compensation assessment and references for small sample machine learning applications.