Soybean Price Trend Forecast Using Deep Learning Techniques Based on Prices and Text Sentiments

Predicting product prices is an essential activity in agricultural value chains. It can improve decision making and revenues for all agents. This chapter explores the use of deep learning techniques for predicting soybeans price trends in Brazil. A long short-term memory neural network (LSTM) forecasts the price signal. A convolutional neural network (CNN) generates a sentiment signal based on the sentiment analysis of news headlines. A multi-layer perceptron (MLP) is also evaluated to generate the sentiment signal, and an ensemble model, composed of both signals, prices and sentiment, is implemented. The four models (LSTM, CNN, and two ensembles with different weights for each signal) are evaluated in terms of their ability to predict the daily price trend. A hyperparameter analysis is conducted for all models, using the mean squared error (MSE) as a metric. Three models obtained the best result (0.60): (i) the LSTM alone; (ii) an ensemble model composed of a simple averaging of the signals; and (iii) an ensemble model composed of 90% price and 10% sentiment. The main findings are: (i) the analysis of the impact of hyperparameters on the models; (ii) the use of dictionaries has not significantly improved the sentiment prediction; (iii) the use of more than 50% of weight in the sentiment signal leads to worse predictions; and (iv) the CNN model provided a better sentiment signal than the MLP model. The benefits and possible uses of the models are discussed. The methodology used can be implemented for other products. Future work is related to improving data sets and implementing econometric models, unsupervised learning, and deep reinforcement learning.