Recurrent neural networks and transformer-based models for multi-step prediction of agricultural commodity prices
Year: 2023
Type: dissertation
Abstract: This dissertation explores the use of Recurrent Neural Networks (RNNs) and Transformerbased models for multi-step prediction of agricultural commodity prices. The study begins with a systematic literature review, providing a comprehensive understanding of the field and identifying gaps in research. Two experimental papers are then presented, focusing on the application of RNNs, specifically Long Short-Term Memory (LSTM) and Bidirectional Long Short-Term Memory Networks (BiLSTM) models, as well as a Transformer-based model for predicting agricultural commodity prices. The commodities of particular interest are soybeans, corn, cattle, and sugar, especially soybeans, the target of the analysis. The study investigates the performance of the Transformer model in forecasting agricultural commodity prices, comparing it with RNNs, both in short and long-term forecasting horizons. The findings highlight the importance of hyperparameter optimization, unbiased model evaluation, and the selection of a suitable forecast horizon that minimizes errors while providing useful information to farmers. Overall, this research contributes to advancing knowledge in the field of RNN models for predicting agricultural commodity prices. It underscores the significance of optimization techniques and the versatility of Transformer models across different domains.
Author: Meloca, Lucas Grogenski
Institution: Universidade Estadual de Maringá
Language: English
Field: Decision Sciences
Domain: Social Sciences
SDG: Zero hunger
Open Access status: green
OpenAlex ID: W7120653588