Forecasting Soybean Futures Prices With Adaptive AI Models

Abstract

The volatility of soybean prices in international futures markets presents significant challenges for stakeholders, including policymakers, producers, and investors. Traditional econometric and time-series models often struggle to capture complex, nonlinear market dynamics, while deep learning models require extensive data and tuning for optimal performance. This study addresses this gap by developing adaptive machine learning models tailored for soybean price forecasting, leveraging historical data from 2020 to 2024, global market reports, and external economic indicators such as exchange rates, crude oil prices, and climate data. Advanced forecasting methods, including Gradient Boosting, Random Forest, ARIMAX, and LSTM, are evaluated alongside a naïve baseline model to benchmark improvements in predictive accuracy. A comparative analysis based on Mean Absolute Percentage Error (MAPE), Mean Absolute Error (MAE), and Root Mean Squared Error (RMSE) identifies Gradient Boosting as the most accurate model, achieving the lowest MAPE of 5.8239%, significantly outperforming deep learning models such as LSTM (MAPE =10.5665%) and ANN (MAPE =26.6723%). Compared to the naïve forecasting model (MAPE =48.0720%), Gradient Boosting reduces error by 87.88%, demonstrating its superior capability in capturing price fluctuations and short-term trends. This study introduces a novel integration of external economic indicators with adaptive feature selection techniques (PCA and L1 Regularization) to improve model interpretability and computational efficiency. Additionally, a rolling update mechanism is implemented, allowing the models to dynamically adapt to evolving market conditions an approach not commonly explored in prior soybean price forecasting studies. These enhancements distinguish this research from existing static forecasting methodologies, improving both accuracy and practical applicability for decision-makers.