Artificial neural networks in freight rate forecasting

Abstract

Reliable freight rate forecasts are essential to stimulate ocean transportation and ensure stakeholder benefits in a highly volatile shipping market. However, compared to traditional time-series approaches, there are few studies using artificial intelligence techniques (e.g. artificial neural networks, ANNs) to forecast shipping freight rates, and fewer still incorporating forward freight agreement (FFA) information for freight rate forecasts. The aim of this paper is to examine the ability of FFAs to improve forecasting accuracy. We use two different dynamic ANN models, NARNET and NARXNET, and we compare their performance for 1, 2, 3 and 6\xa0months ahead. The accuracy of the forecasting models is evaluated with the use of mean squared error (MSE), based on actual secondary data including historical Baltic Panamax Index (BPI) data (available online), and primary data on Baltic forward assessment (BFA) collected from the Baltic Exchange. The experimental results show that, in general, NARXNET outperforms NARNET in all forecast horizons, revealing the importance of the information contained in FFAs in improving forecasting accuracy. Our findings provide better forecasts and insights into the future movements of freight markets and help rationalise chartering decisions.