Forecasting bunker price using deep learning and dimensionality reduction

What the paper says

Purpose Bunker price forecasting is an important task in the shipping industry. Many researchers have contributed to bunker price forecasting using deep learning (DL) models, but have not applied dimensionality reduction techniques (DRTs) to enhance the model accuracy and efficiency. This study proposes a novel hybrid model for forecasting Rotterdam HSFO 380cst prices by combining two DRTs with a deep neural network (DNN). Design/methodology/approach This study combines a mean squared error (MSE) filter and principal component analysis with DL models to produce a six-month-ahead forecast (April–September 2025). Moreover, to assess and compare the quality and accuracy of the models' forecasts, this study uses mean absolute error, MSE, root mean squared error, and mean absolute percentage error (MAPE). In addition, this study uses the Diebold-Mariano and the Harvey, Leybourne, and Newbold tests to test the hypothesis that there is a significant difference between the proposed DL model's forecasts and other DL models. Findings The experimental results confirm that the proposed method enhances the model's efficiency and accuracy. For example, MAPE for the test data decreases significantly from 8.83% (DL model without DRTs) to 5.17% (DL model with DRTs). Originality/value This study combines DRTs with DNNs to build an efficient and robust forecasting model to predict Rotterdam HSFO 380cst prices. The outcomes of this research can be generalised to other complex maritime time series data, such as crude oil and alternative green fuel prices.