Social media-based multi-modal ensemble framework for forecasting soybean futures price

• Construct a sentiment intensity index (ICSI) considering investor concerns. • Build a multi-task autoencoder to convert news with investor concerns into a sentiment index that reflects price changes. • Establish a social media-based multimodal stacking ensemble forecasting framework (SMMEF). • The model shows high accuracy: MAPE is 1.118% for 1-day forecasts and 4.224% for 30-day forecasts. To improve the accuracy of soybean futures price prediction, this paper constructs a sentiment index considering investor concern (ICSI) based on news text, Baidu search volume, and price series, and further develops a multi-modal ensemble prediction model. Firstly, the paper utilizes Baidu search volume as a measure of investor attention. This data is then concatenated with the word embedding vectors of news text, resulting in news texts with investor attention. Then, a multi-task autoencoder is utilized to transform the news texts into representations related to price fluctuations. This result serves as the desired ICSI. Finally, a social media-based multi-modal ensemble forecasting (SMMEF) model is built by integrating the constructed ICSI, statistical variables, and historical price. In the first stage, Long Short-Term Memory and eXtreme Gradient Boosting are used to learn trend features and price fluctuation features respectively in different modalities. In the second stage, Neural basis expansion analysis with exogenous variables (Nbeatsx) is utilized to capture the complementary information between the two modalities for the final prediction. The effectiveness of the proposed model is validated using soybean futures prices as an example. Experimental results show that the proposed ICSI has a strong predictive ability for soybean futures prices, and the proposed SMMEF@Nbeatsx achieves high forecasting accuracy in both short- and medium-term price predictions. When measuring investor sentiment based on news text, investor concern is a key factor influencing the effectiveness of the constructed sentiment variable. Converting news text into representations related to price fluctuations can provide richer and more comprehensive information for prediction models.