A Comparative Study of Traditional Econometric, Machine Learning, and Deep Learning Models for Route-Level Forecasting of the Shanghai Containerized Freight Index: Focusing on the US West Coast, US East Coast, and Europe Routes

Abstract

Using weekly Shanghai Containerized Freight Index (SCFI) data for the US West Coast, US East Coast, and Europe routes (2019–2026; 382 weeks), this study compares traditional econometric methods (ARIMA, SARIMA), machine learning (Random Forest, XGBoost, SVR), and deep learning (LSTM, GRU, Transformer), together with exogenous-variable specifications, under an identical experimental protocol. The analysis covers multiple forecast horizons, Diebold–Mariano tests, regime-based sensitivity analyses, SHAP-based attribution, and forecast combination. No single model dominates across routes. Random Forest is significantly superior for the highly volatile US West Coast route, whereas SARIMA with Fourier terms prevails for the more stable US East Coast and Europe routes. As the horizon lengthens, multivariate models incorporating exogenous variables gain an increasing advantage, although a univariate Transformer remains best at intermediate horizons for the Europe route. Traditional econometric methods overwhelmingly dominate during surge and crash regimes, and the exogenous drivers with the largest model-based (SHAP) contributions differ by route: geopolitical risk for the trans-Pacific routes and economic policy uncertainty for Europe. Weighted forecast combination improves upon the best single model only at the intermediate horizon, where the optimal model tier shifts. These findings suggest that forecasting techniques should be selected differentially according to route-level volatility structures and market regimes, offering practical implications for freight risk management.