Prediction of Baltic Dry Index Based on GRA-BiLSTM Combined Model

The Baltic dry index (BDI) is not only one of the most important indicators of shipping costs but is also an important barometer of global trade and manufacturing sentiment. The BDI is highly volatile and subject to complex factors, which make it difficult to predict. In this paper, a neural network model-based BDI forecasting system was proposed to effectively forecast the BDI. We used the gray relational degree analysis method to select seven factors with higher correlation from 15 factors affecting the variation of BDI index to be used as input indicators for the bi-directional long short-term memory (BiLSTM) model to forecast BDI. From the experimental results, the prediction model proposed in this paper had an excellent prediction effect on the BDI. The mape value of the prediction result was 9.19%. The accuracy was better than the common machine learning models SVR and REG and the neural network model LSTM. In addition, in order to further optimize the prediction performance of the combined model GRA-BiLSTM, this paper introduced the MIV method to conduct an in-depth analysis of the contribution of each variable to the prediction results. Rice price, Shanghai securities composite index and crude oil price were found to be the three most relevant indicators to the prediction accuracy of the model.