ABSTRACT Selecting methods for monthly container throughput forecasting has become more difficult in the post‐COVID period because the literature now offers a wider set of classical, decomposition‐based, machine learning, and deep learning approaches. This study develops a literature‐based Preference Index (PI) to capture method‐family prominence and combines it with rolling‐origin forecast evaluation using monthly throughput data, measured in twenty‐foot equivalent units (TEUs), from the Port of Los Angeles, the Port of Singapore Authority, and Hong Kong Port. The PI combines use, citation influence, and co‐mention embeddedness to shortlist method families rather than to rank them by forecasting superiority. The empirical phase evaluates representative univariate models: Error–Trend–Seasonal models (ETS), seasonal autoregressive integrated moving average models (SARIMA), seasonal‐trend decomposition using Loess (STL)‐based models as representatives of the Ensemble/Hybrid PI family, artificial neural network/multilayer perceptron models (ANN/MLP), and long short‐term memory/recurrent neural network models (LSTM/RNN), across 1‐, 3‐, 6‐, and 12‐month horizons. Accuracy is assessed using mean absolute scaled error (MASE), mean absolute percentage error (MAPE), root mean squared error (RMSE), Diebold–Mariano tests, and Model Confidence Set (MCS) analysis. Results show that classical and decomposition‐based methods remain strong forecasting anchors, but the best specification differs by port: STL + ARIMA performs best for Los Angeles, STL + ETS and SARIMA are strongest for Singapore depending on horizon, while ETS and SARIMA lead for Hong Kong. The lightly specified univariate neural models do not consistently outperform these alternatives. The study offers a transparent screen‐and‐test workflow and helps clarify when classical models may suffice and when learning‐based approaches may require richer information.
Kamal Sanguri (Wed,) studied this question.