A resilient forecasting approach of China’s export container freight index under geopolitical shocks based on a multi-factor screening and multi-model comparison

IntroductionGeopolitical shocks can cause nonlinear and non-stationary fluctuations in the China Containerized Freight Index (CCFI), posing challenges to conventional forecasting methods. This study evaluates the forecasting resilience of different models under both stable market conditions and periods of extreme volatility.MethodsWe constructed a system of 18 potential influencing factors covering four dimensions: macroeconomic conditions, supply and demand, cost factors, and market correlations. A multi-stage screening procedure combining Pearson and Spearman correlation analyses, significance testing, and variance inflation factor diagnostics identified eight core factors: the U.S. Industrial Production Index, U.S. Consumer Price Index, Baltic Dry Index, U.S. dollar interest rate, Brent crude oil price, China’s Consumer Price Index, second-hand containership price index, and new containership orders. ARIMA, Random Forest, XGBoost, and long short-term memory (LSTM) models were compared within a unified framework. The 2010–2025 sample was divided into a calm period (2010–2019) and an extreme-volatility period (2020–2025).ResultsDuring the calm period, LSTM achieved forecasting performance comparable to that of ARIMA and the tree-based models. During the extreme-volatility period, LSTM recorded an MSE of 45,596.30, an MAE of 180.24, an RMSE of 213.53, and a MAPE of 15.98%, with its R2 value being the closest to zero among the four models. The Diebold–Mariano tests showed that LSTM significantly outperformed ARIMA during the extreme-volatility period, whereas its advantages over Random Forest and XGBoost were not consistently significant.DiscussionLSTM demonstrates greater forecasting resilience than the traditional linear model when CCFI dynamics are disrupted by nonlinear and non-stationary geopolitical shocks. Its gating mechanism may facilitate the representation of temporal dependence and post-shock adjustment. However, its superiority over tree-based machine-learning models is not unconditional and may depend on sample division, feature construction, and hyperparameter settings. These findings provide practical support for conflict-aware route adjustment by shipping companies and geopolitical stress testing by government agencies.