Agribusiness time series forecasting using Wavelet neural networks and metaheuristic optimization: An analysis of the soybean sack price and perishable products demand

Brazilian agribusiness is responsible for almost 25% of the country gross domestic product, and companies from this economic sector may have strategies to control their actions in a competitive market. In this way, models to properly predict variations in the price of products and services could be one of the keys to the success in agribusiness. Consistent models are being adopted by companies as part of a decision making process when important choices are based on short or long-term forecasting. This work aims to evaluate Wavelet Neural Networks (WNNs) performance combined with five optimization techniques in order to obtain the best time series forecasting by considering two case studies in the agribusiness sector. The first one adopts the soybean sack price and the second deals with the demand problem of a distinct groups of products from a food company, where nonlinear trends are the main characteristic on both time series. The optimization techniques adopted in this work are: Differential Evolution, Artificial Bee Colony, Glowworm Swarm Optimization, Gravitational Search Algorithm, and Imperialist Competitive Algorithm. Those were evaluated by considering short-term and long-term forecasting, and a prediction horizon of 30 days ahead was considered for the soybean sack price case, while 12 months ahead was selected for the products demand case. The performance of the optimization techniques in training the WNN were compared to the well-established Backpropagation algorithm and Extreme Learning Machine (ELM) assuming accuracy measures. In long-term forecasting, which is considered more difficult than the short-term case due to the error accumulation, the best combinations in terms of precision was reached by distinct methods according to each case, showing the importance of testing different training strategies. This work also showed that the prediction horizon significantly affected the performance of each optimization method in different ways, and the potential of assuming optimization in WNN learning process.